A Progressive Instructional Paradigm for Physics-informed Neural Networks
DOI:
https://doi.org/10.70891/TML.2026.040053Keywords:
PINNs, mathematical modeling, progressive teaching designAbstract
Physics-informed neural networks as an interdisciplinary field combining deep learning and physical modeling, which needs high demands on student ability to transfer knowledge across disciplines. To address the cognitive gap students face when transitioning from traditional mathematical modeling to deep learning methods, this paper develops a three-stage progressive instructional paradigm. Taking the one-dimensional heat conduction equation as an example, teacher guides students through three successive cognitive stage, i.e., mechanistic modeling, data-driven modeling, and physics-informed modeling, corresponding respectively to the finite difference method, pure data-driven MLP fitting, and the PINNs approach. Rather than introducing PINNs directly as a new algorithm, the proposed design organizes their learning as a problem-driven cognitive progression. This can help students understand the paradigm shift from traditional numerical computation to data-driven learning and ultimately to physics-informed integration.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 IFS/ACM Transactions on Machine Learning

This work is licensed under a Creative Commons Attribution 4.0 International License.