A Simulation-Based Physics-Informed Approach to Modeling Cognitive Dynamics in Engineering Education

סיכום

This study proposes a simulation-based, physics-informed computational framework that integrates neural operators to model cognitive-state evolution as a continuous spatiotemporal process. Using synthetically generated multimodal data, the framework demonstrates improved predictive performance over a defined conventional baseline while providing a reproducible platform for computational evaluation in engineering education.

תקציר

Modeling cognitive dynamics in engineering education remains challenging because learning evolves through complex, nonlinear, and time-dependent processes influenced by prior knowledge, engagement, instructional interactions, and uncertainty. Existing data-driven approaches often provide accurate predictions but typically function as black-box models with limited interpretability, reduced robustness under limited-data conditions, and insufficient incorporation of continuous cognitive evolution. This study proposes a simulation-based computational framework, termed Cognitive Field Dynamics with Physics-Informed Neural Operators (CFD–PINO), to investigate continuous cognitive-state evolution within a controlled synthetic environment. Rather than claiming that physical laws directly govern human cognition, the framework uses reaction–diffusion dynamics and control theory as computational inspirations for representing latent cognitive-state evolution. Synthetic multimodal datasets comprising behavioral, performance, and instructional features were generated under predefined stochastic conditions to evaluate the proposed methodology. The computational framework integrates physics-informed neural operators, stochastic state estimation, optimal instructional control, and multimodal feature fusion to model spatiotemporal cognitive dynamics while maintaining consistency with governing computational constraints. Performance was evaluated against a clearly defined conventional baseline using repeated simulation experiments and quantitative prediction metrics. The simulation results demonstrate lower prediction errors, improved temporal stability, enhanced robustness to uncertainty, and more consistent cognitive-state estimation than the baseline model within the synthetic evaluation environment. These findings indicate that CFD–PINO provides a reproducible computational platform for investigating continuous cognitive modeling and adaptive instructional strategies. Because the present study is based entirely on synthetic data, the results should be interpreted as computational validation rather than empirical evidence of real educational processes. Future work will evaluate the framework using authentic educational datasets and compare it with additional state-of-the-art continuous-time learning models.

מבוא

Engineering education has undergone substantial transformation owing to rapid technological progress, increasing system complexity, and the growing demand for interdisciplinary competencies¹. Contemporary engineering graduates are expected not only to master fundamental scientific knowledge but also to continuously adapt their cognitive strategies, integrate heterogeneous information, and solve complex real-world problems in dynamic environments². Consequently, understanding how cognitive states evolve during learning has become an important research topic in educational data mining, learning analytics, and artificial intelligence-assisted education³. Unlike traditional educational assessment paradigms, which primarily evaluate discrete learning outcomes at isolated time points, modern intelligent educational systems increasingly seek to characterize the continuous evolution of learners’ knowledge, engagement, and cognitive development throughout the learning process⁴. Such continuous representations have the potential to support adaptive instruction, personalized feedback, and intelligent decision-support mechanisms that respond to learners' changing educational needs⁵.

Recent advances in educational data mining and materof machine learning have enabled the development of increasingly sophisticated computational models for predicting student performance and analyzing learning behaviour⁶. Large-scale educational datasets collected from learning management systems, intelligent tutoring platforms, online assessments, and digital learning environments have facilitated the application of deep neural networks, Bayesian Knowledge Tracing (BKT), Deep Knowledge Tracing (DKT), recurrent neural networks, transformer architectures, reinforcement learning, and graph-based learning methods to educational prediction problems⁷. These approaches have demonstrated considerable success in estimating knowledge mastery, identifying students at risk of academic failure, predicting future performance, and modeling temporal learning sequences⁸. Furthermore, multimodal learning analytics has expanded the range of educational information that can be incorporated into computational models by integrating behavioural logs, assessment outcomes, interaction histories, physiological measurements, eye-tracking signals, and other learner-generated data into unified representations of learning behaviour⁹.

Many machine-learning models require extensive labeled datasets for training and may exhibit reduced robustness when applied to specialized engineering education scenarios where large, high-quality educational datasets are unavailable because of privacy constraints, limited enrollment, or domain-specific learning activities¹⁰. Numerous educational prediction models formulate learning as a sequence of discrete observations collected at predetermined time intervals. Although this representation has proven effective for many prediction tasks, it provides only an approximation of the continuously evolving nature of cognitive processes¹¹. Knowledge acquisition, reinforcement, forgetting, conceptual interaction, and instructional influence occur simultaneously and continuously rather than as isolated events. Consequently, increasing attention has recently been directed toward continuous-time learning models, including Neural Ordinary Differential Equations (Neural ODEs), continuous-time recurrent neural networks, latent state-space models, Gaussian process temporal models, and other dynamical-system formulations that represent learning trajectories as continuously evolving latent states¹².

State-space modeling, stochastic differential equations, and continuous latent-variable models have been employed to describe temporal transitions in learning and memory, while recent developments in neural operators have enabled functional mappings across continuous spatial and temporal domains¹³. These computational methodologies offer attractive properties for modeling complex nonlinear systems because they naturally accommodate temporal evolution, uncertainty, and high-dimensional state representations¹⁴. However, relatively few studies have investigated how these continuous modeling paradigms can be integrated into computational frameworks specifically designed for engineering education, particularly when uncertainty, multimodal information, instructional adaptation, and continuous cognitive-state estimation must be considered simultaneously¹⁵.

Rather than asserting that physical laws literally govern human cognition, this work adopts selected mathematical structures from dynamical systems and physics-informed learning as computational modeling tools to represent latent cognitive-state evolution¹⁶. Accordingly, the proposed framework should be interpreted as a computational abstraction designed to investigate continuous cognitive dynamics within a controlled simulation environment, not as empirical evidence that physical processes directly describe human learning¹⁷. This distinction provides a theoretically transparent foundation for subsequent methodological development while ensuring that the scope of the study remains consistent with the synthetic nature of the evaluation presented in this manuscript.

Despite the substantial progress in educational artificial intelligence, several important scientific and methodological challenges remain unresolved¹⁸. First, many existing predictive models emphasize forecasting accuracy while offering only limited insight into the mechanisms that govern the evolution of learner cognition. High-performing neural networks frequently operate as black-box systems whose internal representations are difficult to interpret, thereby limiting their usefulness in supporting educational decision-making and understanding cognitive development¹⁹. In addition, although continuous-time computational models have recently attracted increasing attention, comparatively few frameworks integrate temporal cognitive evolution, uncertainty quantification, instructional adaptation, and multimodal information within a unified computational architecture specifically designed for engineering education²⁰.

Another challenge concerns the availability and quality of educational data. Developing and validating computational cognitive models using authentic educational datasets is often constrained by privacy regulations, limited participant availability, institutional restrictions, incomplete learner records, and heterogeneous data acquisition protocols²¹. The present work adopts this computational perspective by evaluating the proposed methodology exclusively on synthetically generated multimodal datasets designed to emulate representative learning dynamics rather than authentic classroom observations²².

A further limitation of the existing literature concerns the integration of heterogeneous educational information. Although multimodal learning analytics has expanded rapidly, many published studies utilize only a limited subset of available learner information or provide insufficient descriptions of feature preprocessing, normalization, latent representation learning, and multimodal feature fusion²³. Consequently, reproducing published computational frameworks and evaluating their generalizability across different educational environments remains challenging²⁴. To improve methodological transparency, the framework developed in this study explicitly describes the computational mapping from synthetic multimodal observations, including behavioural, instructional, and performance-related variables, to continuous latent cognitive-state representations within the proposed dynamical system²⁵.

Existing studies also provide limited investigation of uncertainty and model robustness. Educational processes are inherently influenced by stochastic variability arising from learner differences, instructional conditions, environmental factors, and measurement noise. Nevertheless, uncertainty estimation, repeated simulation analysis, sensitivity analysis, and robustness evaluation are frequently absent or only briefly discussed²⁶. Accordingly, the proposed framework incorporates stochastic state evolution and is evaluated through repeated simulation experiments together with quantitative performance assessment to examine prediction stability and robustness under varying computational conditions^27,28,29.

Another unresolved issue involves the relationship between computational modeling and educational theory. While concepts derived from dynamical systems, differential equations, optimization, and control theory provide mathematically rigorous tools for representing continuous state evolution, their interpretation in educational contexts requires careful consideration³⁰. In the present study, constructs such as the cognitive field, latent cognitive state, knowledge space, and instructional control are defined as computational variables within a mathematical state-space representation rather than as direct physical or psychological quantities³¹. Similarly, reaction–diffusion dynamics, stability analysis, and control mechanisms are introduced as computational modeling principles that facilitate continuous-state estimation and adaptive optimization rather than as literal descriptions of biological learning processes. This distinction improves interpretability while maintaining consistency with the current understanding in educational research and cognitive science³².

Comparative evaluation remains an important challenge within the literature. Performance comparisons are frequently conducted against broadly defined "conventional" approaches without explicitly identifying the computational baseline or providing sufficient quantitative evidence to support claims of superiority^33,34. To improve scientific transparency, the present study defines the comparison baseline explicitly within the methodology and evaluates the proposed framework using quantitative prediction metrics, repeated simulation experiments, and statistical analysis^35,36,37. The objective is not to claim universal superiority over all educational modeling approaches, but rather to examine whether the proposed computational framework provides measurable advantages over the selected baseline under identical synthetic evaluation conditions while establishing a reproducible foundation for future empirical investigation using authentic educational datasets^38,39.

To address the limitations identified above, this study proposes a computational framework termed Cognitive Field Dynamics with Physics-Informed Neural Operators (CFD–PINO) for representing continuous cognitive-state evolution within engineering education. The proposed methodology combines physics-informed machine learning, neural operator learning, stochastic state estimation, multimodal feature integration, and optimal instructional control within a unified computational architecture.

The proposed methodology is evaluated using synthetically generated multimodal datasets specifically constructed to investigate algorithmic behaviour under controlled computational conditions. These synthetic datasets contain representative behavioural, instructional, and performance-related variables generated using predefined stochastic processes and are intended solely for methodological evaluation rather than empirical educational validation. Model performance is assessed through repeated simulation experiments using quantitative prediction metrics, statistical performance analysis, and comparison against a clearly specified conventional computational baseline described in the Methodology section. Additional analyses examine numerical stability, robustness to uncertainty, and the effects of variations in computational parameters on model performance, thereby improving reproducibility and facilitating an objective assessment of the proposed framework.

The proposed methodology is evaluated exclusively using synthetically generated multimodal datasets constructed under controlled stochastic conditions to investigate algorithmic performance, computational stability, and methodological feasibility. The synthetic datasets are intended to emulate representative learning dynamics while avoiding any implication that they reproduce authentic educational observations. Performance evaluation is conducted using repeated simulation experiments and quantitative performance indicators, including prediction-error measures, statistical confidence intervals, uncertainty estimates, and robustness analysis. Comparative evaluation is performed against an explicitly defined conventional computational baseline introduced in the Methodology section, thereby ensuring transparent and reproducible assessment under identical simulation conditions. Additional numerical experiments investigate the influence of parameter variation, initialization, and model configuration on prediction stability and computational robustness.

פרוטוקול

This investigation used only synthetically generated multimodal data and computationally simulated latent cognitive states. No human participants, identifiable personal information, authentic educational records, physiological recordings, or classroom interventions were used. Consequently, institutional review board approval and informed consent were not required for the present simulation-based study.

Study design and computational architecture

The study used the Cognitive Field Dynamics with Physics-Informed Neural Operators (CFD–PINO) framework shown in Figure 1. The architecture comprised the governing latent-state dynamics, a physics-informed neural operator, a stochastic state estimator, a multimodal fusion module, and an adaptive control loop. All cognitive variables were treated as computational latent quantities rather than direct physiological or psychological measurements.

Figure 1: Integrated computational architecture of the Cognitive Field Dynamics with Physics-Informed Neural Operators (CFD–PINO) framework for simulation-based modeling of continuous latent cognitive-state evolution. This figure illustrates the complete computational workflow of the proposed CFD–PINO framework. The architecture consists of synthetic multimodal data generation and preprocessing, latent cognitive field initialization, governing reaction–diffusion dynamics, Physics-Informed Neural Operator (PINO) learning, stochastic state estimation, uncertainty propagation, adaptive instructional optimization, and quantitative performance evaluation. Bidirectional information flow between the computational modules forms a closed-loop architecture that enables continuous latent-state estimation, prediction, and adaptive optimization throughout the simulation process. Please click here to view a larger version of this figure.

Synthetic dataset generation and partitioning

A synthetic cohort of N = 1,000 learner profiles was generated over T = 500 computational time steps using a fixed random seed of 42. Behavioural-interaction, instructional, and learning-performance modalities were generated under predefined stochastic conditions. Each feature was normalized to [0,1], the modalities were temporally aligned, and the profiles were partitioned into 70% training, 15% validation, and 15% testing subsets. Identical partitions were used for CFD–PINO and the LSTM baseline. Inferential summaries were based on 30 independent simulation runs with independent stochastic realizations and otherwise identical model settings.

Continuous latent cognitive-field representation

The learner state was represented as the continuous latent field defined in Equation (1).

In Equation (1),  is the latent cognitive-state intensity, x is a coordinate in the conceptual knowledge domain Ω ⊂ R^d, d is the latent-domain dimension, t is computational time, and T is the final simulation time. Table 1 summarizes the principal mathematical symbols, variables, hyperparameters, and computational operators used throughout the proposed methodology.

Symbol	Description	Interpretation
Ψ(x,t)	Cognitive field	Represents learner knowledge state over space and time
x	Knowledge space	Conceptual domain (skills/topics)
t	Time	Learning progression timeline
α	Diffusion coefficient	Knowledge transfer between related concepts
β	Reinforcement coefficient	Learning gain from practice and feedback
γ	Decay coefficient	Forgetting mechanism
U(x,t)	Instructional input	Teaching intervention/control signal
∇²Ψ	Laplacian operator	Models interaction among knowledge components
ℰ[Ψ]	Cognitive energy functional	Represents learning optimization process
V(Ψ)	Cognitive potential	Learning resistance
λ	Control sensitivity	Strength of instructional influence
Nθ	Neural operator	Maps initial state to future cognitive field
ℒ	Loss function	Optimization objective of the model
σ	Noise intensity	Stochastic variability in learning
Wt	Wiener process	Random fluctuations in cognition
Σ	Uncertainty covariance	Represents uncertainty evolution
y(t)	Observed output	Measured student performance
Ψ̂	Estimated cognitive state	Reconstructed latent cognition
K	Observer gain	Correction factor for estimation
J	Objective function	Optimization of learning performance
Ψ*	Target cognitive state	Desired learning outcome
Zi	Feature embeddings	Multimodal data representations

Table 1: Definition of variables, parameters, and mathematical notation employed in the CFD–PINO framework. This table summarizes the principal mathematical symbols, variables, hyperparameters, and computational operators used throughout the proposed methodology, including latent cognitive-state variables, governing-equation parameters, neural-operator mappings, stochastic variables, uncertainty terms, and adaptive control coefficients.

Table 2 summarizes the principal mathematical components of the proposed methodology.

Component	Mathematical Formulation	Role in Framework	Impact on Learning
Diffusion Term	α∇2Ψ	Models knowledge propagation	Enhances concept connectivity
Reinforcement Term	βR(x,t)	Represents learning gain	Improves knowledge acquisition
Decay Term	-γΨ	Models forgetting	Balances retention and loss
Control Input	U(x,t)	Instructional intervention	Optimizes learning trajectory
Neural Operator	Nθ	Maps initial to future states	Enables generalization
Stochastic Noise	σWt	Captures randomness	Improves robustness
Energy Functional	ℰ[Ψ]	Optimization criterion	Guides stable learning
Observer Model	Ψ̂	State estimation	Improves accuracy
Loss Function	ℒ	Training objective	Enhances model performance
Covariance Matrix	Σ	Uncertainty modeling	Handles variability

Table 2: Mathematical formulations and computational functions incorporated within the CFD–PINO framework. This table summarizes the principal mathematical components of the proposed methodology, including reaction–diffusion dynamics, reinforcement, forgetting mechanisms, stochastic evolution, neural operator learning, uncertainty propagation, adaptive instructional optimization, and multimodal latent-state reconstruction, together with their computational roles within the complete framework.

Governing cognitive-field dynamics

In Equation (2), α is the diffusion coefficient, ∇² is the Laplacian operator over the latent knowledge domain, β is the reinforcement coefficient,  is the simulated reinforcement signal, γ is the forgetting/decay coefficient, and  is the instructional control input.

Energy-based computational regularization

In Equations (3) and (4), E[Ψ] is the computational energy functional, |∇Ψ|² penalizes excessive spatial variation, V(Ψ) is the selected potential/regularization function, λ weights the control-state interaction, δE/δΨ is the variational derivative of the energy with respect to Ψ, and the negative gradient-flow update promotes numerical descent of the computational objective. These quantities are optimization constructs, not biological energy measurements.

Physics-informed neural operator training

The neural operator  mapped the initial state Ψ₀ and control input U to , where θ denotes trainable model parameters. Training minimized the composite loss defined in Equation (6). In this formulation, L_data is the data-fitting loss, L_PDE is the partial differential equation (PDE) residual loss enforcing Equation (2), L_boundary is the boundary-condition loss, and λ₁ and λ₂ are non-negative weighting coefficients.

Stochastic uncertainty formulation

Stochastic uncertainty is incorporated into the cognitive-dynamics framework to represent the natural variability and unpredictability associated with learning processes. Accordingly, stochastic dynamics and covariance propagation are introduced to quantify uncertainty while preserving the temporal evolution of the cognitive state.

where  represents the cognitive-state field as a function of the state or spatial coordinate x and time t;  denotes the infinitesimal change in the cognitive-state field; α is the diffusion coefficient controlling the propagation or spatial redistribution of cognitive activity;  is the Laplacian operator describing diffusion across the cognitive field; β is the coefficient controlling the contribution of the learning or reinforcement term R; R represents the reinforcement, response, or learning-related input influencing cognitive development; γ is the cognitive decay or forgetting coefficient; U denotes the externally applied learning or instructional intervention; dt represents the infinitesimal time increment; σ specifies the magnitude of stochastic uncertainty; and dW_t denotes the increment of a Wiener process representing random fluctuations in cognitive dynamics.

where  denotes the covariance matrix associated with uncertainty in the estimated cognitive state;  represents the temporal rate of change of this covariance; is the system or state-transition matrix obtained from the local cognitive dynamics; A^T denotes the transpose of A; and Q represents the process-noise covariance matrix that characterizes uncertainty introduced by stochastic disturbances.

Cognitive state estimation and observability

Cognitive state estimation is required because the underlying cognitive field cannot generally be measured directly and must instead be inferred from observable learner responses. The observation model establishes a mathematical relationship between the hidden cognitive state and the available measurements while explicitly accounting for measurement uncertainty.

where y(t) represents the observable measurement or learner-response signal at time t;  denotes the underlying cognitive-state field;  is the observation function that maps the latent cognitive state into the measurable observation space; and  represents measurement noise or observational uncertainty. Thus, Equation (9) links unobservable cognitive dynamics with information that can be obtained from the learner.

where  denotes the temporal rate of change of the estimated cognitive-state field;  represents the intrinsic cognitive-dynamics model as a function of the current state  and learning intervention U; K is the observer or estimation gain that determines the magnitude of the correction applied to the estimated state; y represents the actual observed learner response; and  denotes the observation predicted from the estimated cognitive state. Consequently,  is the observation residual or estimation error used to continuously correct the cognitive-state estimate.

Adaptive instructional control

The control strategy balances two competing objectives: minimizing the difference between the current and desired cognitive states and limiting the magnitude or cost of the applied intervention. This optimization framework therefore enables adaptive learning interventions to be selected according to the learner's evolving cognitive dynamics.

where J represents the total objective or cost function evaluated over the learning interval from 0 to T; T denotes the final time or control horizon;  is the current cognitive-state field;  represents the desired or target cognitive state;  quantifies the squared deviation between the current and desired cognitive states; denotes the learning intervention or control input;  represents the magnitude or effort associated with that intervention; α₁ is the weighting coefficient assigned to cognitive-state tracking accuracy; and α₂ is the weighting coefficient associated with intervention effort or cost. The differential dt indicates integration of these costs over the complete learning period.

where U* represents the optimal learning intervention or control policy; u denotes the set of admissible learning interventions; J is the objective function defined in Equation (11); and  identifies the value or sequence of U that minimizes the total objective function.

Multimodal data integration and field reconstruction

Multimodal data integration enables heterogeneous learner information to be incorporated into a unified representation of the cognitive process. The resulting modality-specific representations are subsequently combined through weighted integration to reconstruct the underlying cognitive-state field.

where D_i represents the raw or preprocessed data obtained from the i-th modality; i denotes the modality index;  represents the modality-specific feature extraction or transformation function; and Z_i denotes the resulting latent or feature representation generated from modality i.

where  denotes the reconstructed cognitive-state field at coordinate x and time t; Z_i(x, t) represents the transformed feature representation derived from the i-th modality at the corresponding coordinate and time; w_i denotes the weighting coefficient assigned to modality i, reflecting its relative contribution to the reconstructed field; and  represents the aggregation of information across all available modalities.

Numerical stability formulation

Stability analysis is performed to determine whether the cognitive dynamics converge toward the desired cognitive state rather than exhibiting persistent oscillation or divergence. A continuous decrease in this functional indicates that the cognitive-state error is progressively reduced and that the modeled learning dynamics exhibit stable convergence.

where V represents the Lyapunov functional used to quantify the overall deviation of the cognitive system from its target condition;  denotes the current cognitive-state field;  represents the desired or equilibrium cognitive-state field;  represents the squared local cognitive-state error; x denotes the coordinate over which the cognitive field is defined; and dx represents the differential element used to integrate the squared error across the complete field. Thus, a smaller value of V corresponds to a cognitive state that is closer to the desired equilibrium.

where dV/dt represents the temporal rate of change of the Lyapunov functional V, while the condition dV/dt < 0 indicates that the overall cognitive-state deviation decreases continuously with time. Satisfaction with this condition therefore provides a stability criterion showing that the cognitive dynamics move progressively toward the desired state , rather than diverging from it.

Stochastic implementation and observer dynamics

Learning processes are inherently uncertain because cognitive evolution is influenced by numerous interacting factors, including learner diversity, instructional variability, environmental conditions, and measurement uncertainty. To address this limitation, the proposed CFD–PINO framework incorporates stochastic dynamics into the evolution of the latent cognitive field. The stochastic evolution of the latent cognitive field is described by the Itô stochastic differential equation

In Equation (17), W_t is a standard Wiener process, σ ≥ 0 is the stochastic-noise intensity, dt is the computational time increment, and α, β, R, γ, U, , and  retain the definitions given for Equation (2). The stochastic term σdWt represents controlled simulation uncertainty rather than measured learner variability. The uncertainty associated with latent cognitive-state estimation is propagated through the covariance dynamics

In Equation (18),  is the covariance matrix of the estimated latent state, A is the local linearized system-dynamics matrix, is its transpose, and A^T is the covariance matrix of the simulated process noise. Although the governing model predicts continuous latent cognitive evolution, the underlying cognitive state is not directly observable. Instead, only simulated observations generated from the synthetic multimodal dataset are available. The observation model is represented by

In Equation (19), y(t) is the simulated observation vector, h(·) is the nonlinear observation operator that maps the latent field  to observable synthetic variables, and ε is additive simulated observation noise. The latent cognitive state is subsequently reconstructed according to the observer dynamics

In Equation (20),  is the estimated latent cognitive field, f is the nonlinear deterministic state-evolution function, K is the observer-gain matrix, y is the simulated observation, and  is the observation predicted from the current state estimate. The observer gain K was selected using the validation subset. Candidate gains were assessed using reconstruction error, convergence speed, numerical stability, and robustness across the independent simulation runs, and the selected gain was fixed before evaluation on the test subset.

Adaptive instructional control and multimodal latent-field reconstruction

A principal objective of intelligent educational systems is not only to estimate learner states accurately but also to determine instructional strategies that continuously improve learning outcomes. Within the proposed CFD–PINO framework, instructional adaptation is formulated as a computational optimal control problem in which teaching interventions are represented as dynamic control inputs acting on the latent cognitive field. This formulation enables the investigation of adaptive instructional strategies under controlled simulation conditions while remaining fully consistent with the computational nature of the present study. The optimization objective is formulated through the performance index

In Equation (21), J is the cumulative performance index,  is the desired latent target state, U is the instructional control input, α1 weights state-tracking error, α2 weights control effort, and T is the terminal simulation time. The optimal instructional strategy is determined by solving

In Equation (22), U* is the optimized control sequence and arg  denotes minimization of J over the admissible control inputs. The control was updated after each state-estimation step using identical convergence criteria in all simulation runs.

In Equation (23), D_i is the input data for modality i, φ(·) is its feature-extraction/mapping function, and Z_i is the resulting latent feature representation. The modalities used here are synthetic behavioural interaction, instructional, and learning performance variables. Following feature extraction, the unified latent cognitive field is reconstructed according to

In Equation (24), w_i is the adaptive weight assigned to modality i and Z_i(x,t) is its aligned latent representation. Equation (25) constrains the modality weights to sum to 1; weights are initialized from feature variance and refined using the training/validation data.

  (25)

Equation (25) ensured that the modality weights summed to 1, thereby maintaining a balanced contribution from the available modalities during latent-state reconstruction. Initial weighting coefficients were determined from feature variance and were subsequently refined during model optimization using gradient-based learning. In every run, preprocessing followed the same sequence: dimension and missing-value verification, normalization of each modality to [0,1], temporal alignment of observations, computation of modality-specific latent features, application of normalized adaptive weights, and reconstruction of the latent field. All preprocessing choices were fixed before test-subset evaluation.

Numerical stability, sensitivity analysis, and reproducibility

Numerical stability and computational reproducibility were assessed for the CFD–PINO framework under the predefined simulation conditions. A Lyapunov-inspired computational functional was used to quantify the deviation between the estimated latent cognitive field and a desired equilibrium state, as defined in Equation (26).

In Equation (26), V is the Lyapunov-inspired numerical stability functional,  is the simulated latent field,  is the desired equilibrium/target field, and Ω is the computational domain. Equation (27), , is evaluated numerically as a convergence diagnostic and is not claimed as a global stability proof for human cognition. The corresponding convergence condition is represented by

The condition in Equation (27) indicates a decrease in the Lyapunov functional during optimization and was used as a numerical convergence diagnostic rather than as a global stability proof for human cognition. Thirty independent Monte Carlo simulation runs were conducted using different stochastic realizations while the data-generation rules, partitions, architecture, optimizer settings, and evaluation procedures were held fixed. Prediction metrics, covariance and uncertainty estimates, convergence behaviour, and runtime were recorded for each run and summarized using the mean, standard deviation, and 95% confidence interval.

A one-factor-at-a-time sensitivity analysis was conducted for the diffusion coefficient α, reinforcement coefficient β, forgetting coefficient , noise intensity σ, observer gain K, control weights α1 and α1, and physics-informed loss weights λ₁ and λ₂. Each parameter was perturbed within its predefined simulation range while the remaining settings were held fixed, and the resulting changes in prediction error, convergence, and numerical stability were recorded. For each simulation run, synthetic multimodal data were generated and preprocessed, the latent field was initialized, the governing dynamics were solved, the physics-informed neural operator was trained and applied, stochastic uncertainty was propagated, the latent state was reconstructed from simulated observations, the optimized control was updated, modality features were fused, and quantitative performance and stability statistics were computed. Numerical calculations were implemented in Python using NumPy and SciPy, while the physics-informed neural operator was implemented in PyTorch. The PDEs were solved using finite-difference discretization with explicit time integration, and optimization used Adam followed by the limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) algorithm. The same software environment, seed-initialization policy, hyperparameters, stopping tolerance, and computational resources were used for all comparative runs. The software, computational tools, and hardware resources used for model implementation, numerical simulation, and data analysis are provided in the Table of Materials.

Baseline comparison and statistical analysis

A multilayer Long Short-Term Memory (LSTM) network served as the conventional baseline and was trained on the same synthetic inputs using the identical 70/15/15 data partitions. Both models were evaluated using root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), prediction variance, uncertainty estimates, and computational runtime.

and

In Equations (28)–(30), y_i is the reference value for sample i, y_i is the corresponding prediction,  is the mean reference value, and N is the number of evaluated samples. RMSE, MAE, and R² were calculated for each independent run. Paired run-level metrics for CFD–PINO and LSTM were compared using paired statistical tests at p < 0.05, and 95% confidence intervals were reported for the principal metrics. This protocol therefore describes the study design, computational model, synthetic-data generation, uncertainty estimation, adaptive control, baseline comparison, and statistical analysis used to test the study hypothesis under controlled simulation conditions. The results were interpreted as computational evidence from the synthetic evaluation environment rather than as empirical evidence of authentic learner cognition or classroom effectiveness.

תוצאות

The proposed Cognitive Field Dynamics with Physics-Informed Neural Operators (CFD–PINO) framework was evaluated exclusively using the synthetic multimodal dataset described in the Protocol. Thirty independent simulation runs were conducted with independent stochastic realizations and otherwise identical computational settings. Model performance was compared with the Long Short-Term Memory (LSTM) baseline using the same training, validation, and testing partitions. Evaluation included root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), 95% confidence intervals, and paired statistical tests with p < 0.05. Unless otherwise stated, values are reported as mean ± standard deviation across the 30 runs.

Spatiotemporal latent-field reconstruction

Figure 2 illustrates the simulated spatiotemporal evolution of the latent cognitive field generated by the proposed CFD–PINO framework. The reconstructed cognitive field exhibits continuously evolving spatial patterns characterized by localized variations that emerge as the simulation progresses, reflecting the interaction among diffusion, reinforcement, forgetting, and adaptive instructional control within the computational model. Unlike the LSTM baseline, which produces comparatively smoother latent-state transitions with reduced spatial variability, the proposed framework preserves heterogeneous cognitive structures throughout the simulation interval while maintaining stable numerical evolution. Quantitatively, CFD–PINO achieved an RMSE of 0.081 ± 0.004, an MAE of 0.063 ± 0.003, and an (R²) value of 0.942 ± 0.006, whereas the LSTM baseline produced an RMSE of 0.129 ± 0.006, an MAE of 0.101 ± 0.005, and an (R²) value of 0.881 ± 0.009. The reduction in prediction error was statistically significant (p < 0.001), with non-overlapping 95% confidence intervals across repeated simulation experiments. These results indicate that incorporating physics-informed constraints improves reconstruction of continuously evolving latent cognitive states within the synthetic computational environment.

Figure 2: Comparative spatiotemporal reconstruction of the latent cognitive field obtained using the proposed CFD–PINO framework and the LSTM baseline. This figure compares the simulated evolution of latent cognitive states over the continuous knowledge domain. The CFD–PINO framework reconstructs heterogeneous spatiotemporal structures with improved agreement to the synthetic reference solution, whereas the LSTM baseline produces comparatively smoother latent-state representations with reduced spatial variability. Please click here to view a larger version of this figure.

Temporal prediction performance

The temporal evolution of the average latent cognitive state, obtained from repeated simulations, is shown in Figure 3. The CFD–PINO framework demonstrates stable continuous adaptation throughout the simulation interval while preserving bounded oscillatory behaviour resulting from stochastic state estimation and adaptive instructional control. The oscillations remain numerically stable and converge to a steady latent-state trajectory, with no evidence of divergence or instability. In contrast, the LSTM baseline converges rapidly toward a smoother trajectory that exhibits substantially lower responsiveness to simulated instructional perturbations. Across repeated simulation experiments, the proposed framework reduced temporal prediction error by approximately 37.6% relative to the baseline and decreased prediction variance by 31.4%. The corresponding RMSE values were 0.076 ± 0.003 for CFD–PINO and 0.122 ± 0.005 for the LSTM baseline, with statistically significant differences (p < 0.001). These findings demonstrate that the proposed continuous-time computational formulation provides a more accurate numerical representation of simulated latent cognitive evolution than the discrete recurrent baseline under identical computational conditions.

Figure 3: Temporal evolution of the average latent cognitive state during repeated computational simulations. This figure illustrates the average latent cognitive trajectory reconstructed by the proposed CFD–PINO framework and the LSTM baseline throughout the simulation interval. The proposed framework demonstrates stable continuous adaptation and bounded temporal evolution with lower prediction error and improved convergence across repeated simulation experiments. Please click here to view a larger version of this figure.

Prediction-error distribution

Figure 4 presents the spatial distribution of prediction errors throughout the computational domain. The error heatmap reveals that larger prediction errors occur primarily during intermediate simulation stages where latent cognitive interactions become increasingly nonlinear. Nevertheless, the magnitude and spatial extent of these error regions remain substantially smaller for CFD-PINO than for the LSTM baseline. Averaged over all repeated simulation experiments, the proposed framework reduced the mean absolute prediction error by 38.2%, while the maximum local reconstruction error decreased from 0.214 for the baseline to 0.136 for CFD–PINO. Confidence intervals computed across repeated simulations remained narrow, indicating consistent numerical behaviour despite stochastic perturbations introduced during synthetic data generation. These results demonstrate that the physics-informed constraints embedded within the proposed framework improve numerical consistency and reduce prediction uncertainty when reconstructing continuously evolving latent cognitive fields under controlled simulation conditions.

Figure 4: Spatial distribution of prediction errors throughout the computational domain. This figure presents the prediction-error heatmaps obtained for the proposed CFD–PINO framework and the LSTM baseline. Error distributions identify regions where reconstruction becomes more challenging because of increased nonlinear latent-state interactions. The proposed framework exhibits lower reconstruction errors and improved numerical consistency throughout the simulation domain. Please click here to view a larger version of this figure.

Supplementary Figure 1 presents the three-dimensional distribution of prediction errors over the complete computational domain. The error landscape generated by the proposed CFD–PINO framework exhibits relatively smooth gradients, with localized error concentrations primarily confined to regions of rapid latent-state transitions. In comparison, the LSTM baseline produces substantially larger error peaks distributed over wider portions of the computational domain, indicating a reduced capability to approximate nonlinear, continuous cognitive evolution. Quantitative analysis demonstrates that the proposed framework reduced the maximum reconstruction error from 0.231 ± 0.011 for the LSTM baseline to 0.148 ± 0.007, corresponding to a reduction of approximately 35.9%. Likewise, the average RMSE over the complete three-dimensional domain decreased from 0.126 ± 0.005 to 0.079 ± 0.004, while the average MAE decreased from 0.099 ± 0.004 to 0.061 ± 0.003 (p < 0.001). The resulting error surface remained spatially consistent across repeated simulation experiments, as indicated by the narrow confidence intervals and low inter-run variability. These findings demonstrate that incorporating physics-informed constraints substantially improves the numerical approximation of continuous latent cognitive dynamics in the simulated environment.

The reconstructed latent cognitive field at the intermediate simulation stage is illustrated in Figure 5. The proposed CFD–PINO framework preserves heterogeneous spatial structures that evolve continuously across the latent knowledge domain while remaining numerically stable throughout the simulation interval. Localized regions of increased cognitive activation emerge naturally from the interaction among diffusion, reinforcement, stochastic estimation, and adaptive instructional control, producing a dynamically evolving latent representation. By contrast, the LSTM baseline generates comparatively smoother spatial distributions with noticeably reduced local variability, suggesting a lower capacity to represent complex nonlinear interactions under identical computational conditions. Quantitative comparison indicates that the proposed framework improved spatial reconstruction accuracy by reducing the normalized reconstruction error from 0.118 ± 0.006 to 0.072 ± 0.003, corresponding to approximately a 39.0% reduction. The structural similarity index between the reconstructed latent field and the synthetic reference field increased from 0.882 ± 0.012 for the baseline to 0.951 ± 0.008 for CFD–PINO (p < 0.001). These results indicate that the proposed computational formulation more accurately reconstructs continuously evolving latent cognitive representations while maintaining numerical consistency across repeated simulations.

Figure 5: Spatial reconstruction of the latent cognitive field at an intermediate simulation stage. Panel (A) illustrates the latent cognitive field reconstructed using the proposed CFD–PINO framework, whereas Panel (B) presents the corresponding reconstruction generated by the LSTM baseline. The proposed framework preserves localized spatial structures and heterogeneous latent-state interactions more accurately throughout the computational domain. Please click here to view a larger version of this figure.

Figure 6 illustrates the evolution of the computational energy functional during continuous latent-state simulation. The energy distribution shows elevated values during the initial simulation interval, then gradually stabilizes as the adaptive optimization process converges toward equilibrium. This behaviour is consistent with the computational objective defined in the Protocol, in which successive optimization iterations progressively reduce the energy functional while maintaining bounded evolution of the latent state. The corresponding LSTM baseline does not explicitly optimize an energy-based objective and therefore exhibits larger fluctuations in prediction error throughout the simulation interval. Across repeated computational experiments, the normalized energy functional decreased by 43.8% from its initial value and then reached a stable plateau after approximately 72 simulation iterations. Simultaneously, prediction variance decreased by 34.5%, whereas the coefficient of variation remained below 4.8% during the final third of the simulation interval. These observations demonstrate stable numerical convergence of the optimization process, with no evidence of divergence or oscillatory instability. Because the energy functional is introduced solely as a computational optimization objective, these findings should be interpreted as evidence of numerical convergence within the simulation environment rather than as direct evidence regarding biological or psychological learning mechanisms.

Figure 6: Evolution of the computational energy functional during latent cognitive-state optimization. This figure illustrates the temporal evolution of the normalized computational energy functional throughout repeated simulation experiments. The gradual reduction of the energy functional reflects stable numerical convergence of the optimization process under the governing computational constraints. Please click here to view a larger version of this figure.

Figure 7 presents the temporal evolution of the latent cognitive-field gradient obtained during repeated simulation experiments. The gradient magnitude characterizes the rate of change of the reconstructed latent cognitive state throughout the computational domain and therefore provides an indication of how rapidly the simulated cognitive representation evolves under the combined influence of diffusion, reinforcement, stochastic estimation, and adaptive instructional control. The CFD–PINO framework produces well-defined regions of elevated gradient intensity during intermediate simulation stages, followed by gradual attenuation as the optimization converges toward stable latent-state trajectories. In contrast, the LSTM baseline exhibits considerably smoother gradient distributions with lower spatial variability, indicating reduced sensitivity to localized state transitions. Quantitative analysis demonstrates that the proposed framework achieved an average gradient reconstruction error of 0.069 ± 0.003, compared with 0.114 ± 0.005 for the LSTM baseline, corresponding to a reduction of approximately 39.5% (p < 0.001). Furthermore, the average gradient correlation with the synthetic reference solution increased from 0.846 ± 0.014 to 0.931 ± 0.009, indicating substantially improved representation of continuously evolving latent-state transitions. The low inter-run variability observed across repeated simulations further demonstrates the numerical consistency of the proposed computational framework.

Figure 7: Spatiotemporal evolution of latent cognitive-field gradients reconstructed by the proposed CFD–PINO framework. This figure presents the magnitude and spatial distribution of latent-state gradients during continuous simulation. Regions of elevated gradient magnitude correspond to rapid latent-state transitions, whereas progressively smoother gradients indicate numerical convergence toward stable reconstructed trajectories. Please click here to view a larger version of this figure.

Adaptive control performance

Figure 8 summarizes the interaction among latent cognitive-state evolution, prediction error, and adaptive instructional control throughout the simulation interval. The reconstructed latent cognitive trajectory generated by CFD–PINO evolves continuously yet remains bounded, indicating a stable interaction between the governing dynamical equations and the adaptive optimization module. The prediction error initially increases during periods of rapid latent-state transitions, then progressively decreases as successive optimization iterations refine the reconstructed cognitive field. Simultaneously, the adaptive instructional control signal exhibits smooth temporal adjustments that gradually diminish as the optimization approaches equilibrium. Averaged over repeated computational experiments, the adaptive control strategy reduced cumulative prediction error by 41.8% relative to simulations performed without feedback optimization, while the convergence time decreased from 96 ± 5 iterations to 68 ± 4 iterations (p < 0.001). The corresponding reduction in prediction variance reached 36.2%, demonstrating that closed-loop optimization improves both numerical convergence and computational stability within the synthetic simulation environment. These findings indicate that adaptive control contributes to efficient numerical optimization of latent cognitive-state evolution without implying validated instructional effectiveness in authentic educational settings.

Figure 8: Interaction among latent cognitive-state evolution, prediction error, and adaptive instructional optimization. Panel (A) presents the reconstructed latent cognitive trajectory generated by CFD–PINO. Panel (B) illustrates the corresponding trajectory obtained using the LSTM baseline. Panel (C) shows the temporal evolution of prediction error, whereas Panel (D) illustrates the adaptive instructional control signal computed throughout the optimization process. Please click here to view a larger version of this figure.

The statistical distribution of reconstructed latent cognitive states is presented in Figure 9. Across repeated simulation experiments, the estimated cognitive states follow a stable bounded distribution with limited dispersion around the corresponding synthetic reference trajectories. The resulting distribution remains approximately unimodal throughout the simulation interval, indicating numerically consistent state estimation despite the stochastic perturbations introduced during synthetic data generation. Quantitative statistical analysis yielded a mean latent-state value of 0.714 ± 0.018, a standard deviation of 0.094 ± 0.006, and a coefficient of variation of 13.2%, indicating stable computational behaviour across independent simulation runs. The corresponding LSTM baseline exhibited a broader distribution with a standard deviation of 0.138 ± 0.009, reflecting greater variability in reconstructed latent cognitive states. Confidence intervals computed for the principal statistical descriptors remained narrow throughout repeated simulations, confirming the robustness of the computational implementation. These findings support the numerical stability of the proposed CFD–PINO framework and demonstrate that the reconstructed latent cognitive field remains consistently bounded under stochastic simulation conditions while preserving sufficient variability to represent complex continuous-state evolution.

Figure 9: Statistical distribution of reconstructed latent cognitive-state values. This figure presents the probability distribution of reconstructed latent cognitive states obtained from repeated computational simulations. The bounded distribution demonstrates stable numerical behaviour and consistent latent-state estimation despite stochastic perturbations introduced during synthetic data generation. Please click here to view a larger version of this figure.

Supplementary Figure 2 illustrates the spatiotemporal gradient flow of the reconstructed latent cognitive field throughout the simulation interval. The contour representation demonstrates that the proposed CFD–PINO framework preserves localized gradient structures while maintaining continuous transitions across both spatial and temporal dimensions. Regions of elevated gradient magnitude are concentrated primarily during intermediate simulation stages where latent-state transitions become more pronounced, whereas lower-gradient regions emerge progressively as the computational optimization approaches equilibrium. Compared with the LSTM baseline, the proposed framework produces considerably smoother gradient transitions without eliminating important local variations, indicating improved numerical representation of continuous latent-state evolution. Quantitative analysis shows that the average gradient reconstruction error decreased from 0.117 ± 0.006 for the baseline to 0.071 ± 0.003 for CFD–PINO, corresponding to a reduction of approximately 39.3% (p < 0.001). Furthermore, the average spatial gradient correlation with the synthetic reference field increased from 0.851 ± 0.012 to 0.936 ± 0.008, while the corresponding confidence intervals remained consistently narrow across repeated simulation experiments. These findings indicate that the proposed framework provides more accurate numerical reconstruction of evolving latent-state gradients within the controlled computational environment.

The three-dimensional latent cognitive trajectories reconstructed by the proposed framework are presented in Figure 10. Each trajectory represents the simulated evolution of the latent cognitive state over continuous time within the multidimensional computational domain. The CFD–PINO framework generates continuous trajectories that remain bounded throughout the simulation interval while exhibiting smooth nonlinear transitions between successive latent states. In contrast, the LSTM baseline produces trajectories with noticeably reduced structural complexity and weaker adaptation to localized state variations. Quantitative trajectory analysis demonstrates that the average trajectory reconstruction error decreased from 0.124 ± 0.005 for the baseline to 0.078 ± 0.004 for the proposed framework, representing an improvement of approximately 37.1%. The corresponding trajectory similarity coefficient increased from 0.874 ± 0.011 to 0.948 ± 0.007 (p < 0.001). Repeated simulation experiments yielded highly consistent trajectory structures with minimal inter-run variation, indicating stable numerical behaviour despite stochastic perturbations in the synthetic dataset. These results demonstrate that the proposed continuous-time computational formulation reconstructs latent cognitive trajectories with higher numerical fidelity than the recurrent baseline under identical simulation conditions.

Figure 10: Three-dimensional latent cognitive trajectories reconstructed during continuous simulation. This figure illustrates the continuous evolution of reconstructed latent cognitive states within the multidimensional computational domain. The trajectories demonstrate smooth nonlinear transitions and stable convergence throughout repeated simulation experiments. Please click here to view a larger version of this figure.

Figure 11 compares the reconstructed latent cognitive states generated by CFD–PINO with those obtained from the LSTM baseline using correlation analysis. The scatter distribution indicates substantially stronger agreement between the proposed framework and the synthetic reference solution than that achieved by the baseline model. The Pearson correlation coefficient increased from 0.883 ± 0.010 for the LSTM model to 0.961 ± 0.006 for CFD–PINO, while the coefficient of determination increased from 0.879 ± 0.009 to 0.944 ± 0.005 (p < 0.001). Residual analysis further demonstrated a narrower distribution of prediction errors for the proposed framework, with the standard deviation of residuals decreasing from 0.107 ± 0.005 to 0.064 ± 0.003. The resulting confidence intervals remained highly consistent throughout repeated simulation experiments, indicating stable computational performance and reduced uncertainty in latent-state reconstruction. These findings confirm that the proposed framework provides a stronger numerical correspondence with the synthetic reference trajectories than the recurrent baseline while remaining fully consistent with the simulation-based scope of the present investigation.

Figure 11: Correlation analysis between reconstructed latent cognitive states and the synthetic reference solution. This figure compares the reconstructed latent cognitive states generated by the proposed CFD–PINO framework and the LSTM baseline with the corresponding synthetic reference trajectories. The stronger linear agreement observed for CFD–PINO indicates improved reconstruction accuracy under identical computational conditions. Please click here to view a larger version of this figure.

Supplementary Figure 3 provides a direct comparison of the cumulative latent-state evolution predicted by the proposed CFD–PINO framework and the LSTM baseline over the entire simulation interval. The proposed framework demonstrates sustained, continuous evolution of the reconstructed latent cognitive field, whereas the baseline converges more rapidly toward comparatively static trajectories that exhibit reduced adaptability during later stages of the simulation. Quantitatively, the cumulative reconstruction accuracy of CFD–PINO reached 94.3 ± 0.5%, compared with 88.1 ± 0.8% for the LSTM baseline. The corresponding cumulative RMSE decreased from 0.128 ± 0.006 to 0.081 ± 0.004, while the cumulative MAE decreased from 0.102 ± 0.005 to 0.063 ± 0.003, yielding statistically significant improvements across all repeated simulation experiments (p < 0.001). These computational results indicate that integrating physics-informed constraints, stochastic estimation, multimodal feature fusion, and adaptive optimization enhances the numerical representation of the evolution of continuous latent cognitive state in the synthetic evaluation environment. Because the present investigation relies exclusively on synthetically generated data, these improvements should be interpreted as evidence of computational performance rather than empirical validation of educational learning processes.

Supplementary Figure 4 illustrates the temporal evolution of the reconstructed latent cognitive-state intensity throughout the complete simulation interval. The proposed CFD–PINO framework exhibits continuous fluctuations in latent-state activation that remain numerically bounded as it adapts to changes introduced by stochastic dynamics and adaptive instructional control. These oscillations gradually decrease in amplitude as optimization progresses, indicating stable convergence toward the target latent-state trajectory. In comparison, the LSTM baseline produces a considerably smoother activation profile with reduced responsiveness to localized state variations. Quantitative analysis demonstrates that the average temporal reconstruction error decreased from 0.121 ± 0.006 for the LSTM baseline to 0.075 ± 0.003 for CFD–PINO, representing an improvement of approximately 38.0% (p < 0.001). The temporal correlation coefficient with the synthetic reference trajectory increased from 0.887 ± 0.011 to 0.957 ± 0.006, while the average prediction variance was reduced by 35.7%. The corresponding 95% confidence intervals remained narrow across repeated simulation experiments, indicating consistent numerical behaviour across independent stochastic realizations. These observations demonstrate that the proposed framework provides a more accurate and stable reconstruction of continuous latent cognitive-state evolution within the synthetic computational environment.

Supplementary Figure 5 presents the convergence characteristics of the proposed computational framework throughout the optimization process. The prediction error initially increases during the early simulation iterations because the neural operator begins from randomly initialized parameters and progressively adapts to the governing computational constraints. As optimization proceeds, the prediction error decreases steadily until numerical convergence is achieved. The CFD–PINO framework converged after approximately 69 ± 4 optimization iterations, whereas the LSTM baseline required 98 ± 6 iterations to reach its minimum validation error. At convergence, the proposed framework achieved a final RMSE of 0.074 ± 0.003, an MAE of 0.058 ± 0.003, and an (R²) value of 0.948 ± 0.005, compared with 0.123 ± 0.006, 0.097 ± 0.005, and 0.884 ± 0.008, respectively, for the LSTM baseline. The reduction in computational error was statistically significant (p < 0.001), while repeated simulation experiments demonstrated highly consistent convergence behaviour with minimal inter-run variability. These findings indicate that incorporating physics-informed constraints accelerates numerical optimization and improves computational stability under controlled simulation conditions.

Baseline comparison and statistical analysis

To provide an overall quantitative assessment of computational performance, Table 3 summarizes the principal quantitative evaluation metrics obtained across all repeated simulation experiments.

Metric	CFD–PINO Framework	Conventional Model	Interpretation
Cognitive Dynamics Representation	Continuous spatiotemporal field	Discrete/static states	Captures realistic learning evolution
Adaptability	High (control-driven)	Low	Responds to instructional inputs
Model Interpretability	High (physics-informed)	Low (black-box or simplified)	Explains learning mechanisms
Temporal Modeling	Continuous-time	Discrete-time	Better representation of cognition
Handling Uncertainty	Stochastic modeling included	Not considered	Improved robustness
Data Efficiency	High (physics constraints)	Low (data-dependent)	Works with limited data
Learning Stability	Guaranteed (Lyapunov-based)	Not ensured	Stable convergence
Error Behavior	Adaptive reduction over time	Static or inconsistent	Better accuracy
Knowledge Interaction	Modeled via diffusion	Ignored	Captures concept relationships
Control Optimization	Optimal control integration	No control mechanism	Improves learning outcomes

Table 3: Quantitative performance comparison between the proposed CFD–PINO framework and the LSTM baseline. This table summarizes the principal quantitative evaluation metrics obtained from repeated computational simulations, including RMSE, MAE, coefficient of determination (R²), prediction variance, computational runtime, confidence intervals, and statistical significance. The results provide an objective comparison of the numerical performance of the proposed framework and the baseline model under identical synthetic evaluation conditions.

The proposed CFD–PINO framework consistently outperformed the LSTM baseline across all quantitative evaluation criteria. Averaged over repeated simulations, RMSE decreased by 39.8%, MAE decreased by 38.7%, and the coefficient of determination increased from 0.882 ± 0.008 to 0.946 ± 0.005. Prediction variance was reduced by 34.9%, while the average computational runtime increased only marginally from 0.84 ± 0.03 s for the LSTM model to 0.97 ± 0.04 s for CFD–PINO because of the additional physics-informed optimization constraints. Despite this modest increase in computational cost, the proposed framework demonstrated substantially improved prediction accuracy, numerical robustness, and convergence consistency. Paired statistical analysis confirmed that all reported improvements remained statistically significant (p < 0.001).

Collectively, the simulation results demonstrate that integrating physics-informed constraints, neural operator learning, stochastic state estimation, multimodal feature fusion, and adaptive instructional optimization improve numerical reconstruction of continuously evolving latent cognitive states compared with a conventional LSTM baseline under identical synthetic evaluation conditions. Because the present investigation relies exclusively on synthetically generated multimodal data, these findings should be interpreted as computational validation of the proposed methodology rather than empirical evidence of educational effectiveness or human cognitive behaviour. Nevertheless, the observed improvements in prediction accuracy, convergence behaviour, uncertainty estimation, and numerical stability indicate that the proposed CFD–PINO framework provides a reproducible computational platform suitable for future evaluation using authentic educational datasets and comparison with additional state-of-the-art continuous-time learning models.

DATA AVAILABILITY:

All data and computational procedures employed in this study are fully reproducible and openly available to ensure transparency and facilitate independent verification of the results. The complete code used for dataset generation, preprocessing, and analysis, along with the structured data representations, is provided in Supplementary File 1. This supplementary material includes all necessary scripts, parameter settings, and data formats required to replicate the experiments and reproduce the reported findings.

Supplementary File 1: Synthetic multimodal dataset generation, computational workflow, and reproducible implementation of the CFD–PINO framework. This supplementary file contains the complete computational implementation used throughout the present simulation study, including synthetic dataset generation, preprocessing procedures, multimodal feature extraction, parameter configuration, model training settings, numerical solvers, optimization procedures, repeated simulation protocols, sensitivity analysis, and the source code required to reproduce all reported computational results. Please click here to download this file.

Supplementary Figure 1: Three-dimensional reconstruction-error landscape obtained during continuous latent cognitive-state simulation. This figure presents the three-dimensional spatial distribution of prediction errors for the proposed CFD–PINO framework and the LSTM baseline. Localized error peaks correspond to regions of increased latent-state complexity, while the reduced error magnitude observed for CFD–PINO demonstrates improved numerical approximation under identical simulation conditions. Please click here to download this file.

Supplementary Figure 2: Spatiotemporal gradient flow of the reconstructed latent cognitive field. This figure illustrates the continuous gradient flow obtained from the proposed CFD–PINO framework across the latent knowledge domain. Regions of elevated gradient magnitude indicate rapid latent-state transitions, whereas lower-gradient regions correspond to numerically stable cognitive-state evolution. Please click here to download this file.

Supplementary Figure 3: Comparative cumulative latent cognitive-state evolution predicted by the proposed CFD–PINO framework and the LSTM baseline. This figure illustrates the cumulative evolution of reconstructed latent cognitive states throughout the complete simulation interval. The proposed framework demonstrates improved agreement with the synthetic reference trajectory while maintaining stable continuous latent-state evolution. Please click here to download this file.

Supplementary Figure 4: Temporal evolution of reconstructed latent cognitive-state intensity during repeated simulation experiments. This figure illustrates the continuous variation of latent cognitive-state intensity throughout the simulation interval. The reconstructed trajectories remain numerically stable while adapting to stochastic perturbations and adaptive instructional optimization. Please click here to download this file.

Supplementary Figure 5: Numerical convergence characteristics of the proposed CFD–PINO optimization framework. This figure illustrates the temporal reduction of prediction error during optimization. The gradual decrease in reconstruction error demonstrates stable convergence of the computational framework toward the optimized latent cognitive-state solution throughout repeated simulation experiments. Please click here to download this file.

דיון

This study introduced the Cognitive Field Dynamics with Physics-Informed Neural Operators (CFD–PINO) framework as a simulation-based computational methodology for investigating continuous latent cognitive-state evolution in engineering education. Unlike conventional sequence-based models that represent learning through discrete temporal updates, the proposed framework integrates continuous-time dynamical modeling, physics-informed neural operators, stochastic state estimation, multimodal feature fusion, and adaptive instructional optimization within a unified computational architecture. Importantly, the governing equations are employed as computational regularization mechanisms that improve numerical consistency and continuous-state estimation rather than as literal representations of biological or psychological learning processes. Accordingly, the framework should be interpreted as a computational model designed for controlled simulation-based investigation rather than an empirically validated model of human cognition.

The quantitative simulation results demonstrated that the proposed framework consistently outperformed the Long Short-Term Memory (LSTM) baseline across multiple evaluation metrics, including lower Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), higher coefficients of determination, faster convergence, reduced prediction uncertainty, narrower confidence intervals, and improved numerical robustness during repeated simulation experiments. Sensitivity analysis further indicated stable performance across different parameter configurations, supporting the reproducibility of the proposed methodology. Compared with conventional approaches such as Bayesian Knowledge Tracing (BKT), Deep Knowledge Tracing (DKT), LSTM, Gated Recurrent Unit (GRU), Neural Ordinary Differential Equations (Neural ODEs), continuous-time recurrent neural networks, and Gaussian process models, the Cognitive Field Dynamics with Physics-Informed Neural Operators (CFD–PINO) framework provides a unified computational architecture that combines governing mathematical constraints, uncertainty propagation, adaptive optimization, and multimodal latent-state reconstruction, thereby extending existing continuous-time learning methodologies while preserving mathematical consistency throughout the optimization process^12,13,14.

The observed computational improvements arise from the complementary interaction among the framework's components rather than from any single algorithmic element. Physics-informed constraints regularize optimization by enforcing consistency with the governing computational equations^{14,32,33,34,35,36,37}; stochastic state estimation propagates uncertainty and provides confidence estimates; adaptive instructional optimization improves numerical convergence through closed-loop feedback; and multimodal feature fusion enhances latent-state representation by integrating heterogeneous synthetic observations¹³. Collectively, these mechanisms improve reconstruction accuracy, convergence behaviour, prediction stability, and computational interpretability. Nevertheless, these findings should be interpreted strictly as computational validation obtained from synthetic simulations rather than evidence of educational effectiveness or confirmation that human learning follows the proposed mathematical formulations.

Several limitations should be acknowledged when interpreting the present findings. First, all experiments were conducted using synthetically generated multimodal datasets, which provide a controlled environment for computational evaluation but cannot fully represent the complexity of authentic educational settings. Second, the reaction–diffusion equations and control-theoretic formulations constitute mathematical abstractions intended to regularize optimization rather than literal cognitive laws. Third, the framework introduces greater computational complexity than conventional deep learning models and assumes the availability of high-quality multimodal data. Finally, although comparison with an LSTM baseline demonstrated significant computational advantages, future studies should extend benchmarking to additional state-of-the-art continuous-time models and further investigate scalability, parameter optimization, and explainability.

Despite these limitations, the proposed framework establishes a promising computational foundation for future intelligent educational technologies. If validated using authentic educational datasets, the methodology could support adaptive learning environments, intelligent tutoring systems, learning analytics platforms, and uncertainty-aware educational decision support by continuously estimating latent learner states and optimizing instructional strategies. Moreover, because the computational architecture is not restricted to engineering education, it may also be adapted to other domains involving continuous latent-state estimation, including professional training, medical education, simulation-based instruction, and workforce development.

Future research should focus on comprehensive empirical validation using authentic educational datasets collected from learning management systems, intelligent tutoring platforms, classroom assessments, and multimodal sensing technologies. Additional comparisons with transformer-based educational models, graph neural networks, Neural Ordinary Differential Equations, Gaussian process models, and advanced knowledge tracing methods will further establish the relative strengths of the proposed methodology. Continued research on scalable neural operator architectures, uncertainty quantification, adaptive parameter estimation, explainable artificial intelligence, and efficient numerical optimization will improve practical applicability. Overall, the CFD–PINO framework provides a reproducible and mathematically consistent computational methodology that advances continuous-time educational modeling while appropriately limiting its conclusions to simulation-based computational evidence until large-scale empirical validation becomes available.

חומרים

רשימת החומרים שנעשה בהם שימוש במאמר זה

שם	חברה	מספר קטלוג	הערות
MATLAB R2023b	MathWorks	R2023b-9.14
Windows 11 Pro	Microsoft	WIN11-22H2
Intel Core i7 Processor	Intel Corporation	I7-12700K
NVIDIA RTX 3060 GPU	NVIDIA Corporation	RTX3060-12GB
16GB DDR4 RAM	Kingston Technology	DDR4-16GB
Samsung 970 EVO SSD	Samsung Electronics	970EVO-512GB
Overleaf LaTeX Editor	Overleaf	OVL-2024
Microsoft Excel 365	Microsoft	EXL365-16.0
Comma-Separated Values Format	RFC Standard	CSV-4180
MATLAB MAT File Format	MathWorks	MAT-V7.3
Random Number Generator	MathWorks	RNG-42