Physics-Transfer Learning: A Framework to Address the Accuracy-Performance Dilemma in Modeling Complexity Problems in Engineering Sciences

Yingjie Zhao, Zhiping Xu

Keywords: Physics-Transfer Learning; Accuracy-Performance Dilemma; Engineering Sciences; Complexity; Materials Strength; Brain Development

Abstract: The development of theoretical sciences traditionally adheres to an observation-assumption-model paradigm, which is effective in simple systems but challenged by the `curse of complexity’ in modern engineering sciences. Advancements in artificial intelligence (AI) and machine learning (ML) offer a data-driven alternative, capable of interpolating and extrapolating scientific inference where direct solutions are intractable. Moreover, feature engineering in ML resembles dimensional analysis in classical physics, suggesting that data-driven ML methods could potentially extract new physics behind complex data. Here we propose a physics-transfer (PT) learning framework to learn physics across digital models of varying fidelities and complexities, which addresses the accuracy-performance dilemma in understanding representative multiscale problems. The capability of our approach is showcased through screening metallic alloys by their strengths and predicting the morphological development of brains. The physics of crystal plasticity is learned from low-fidelity molecular dynamics simulation and the model is then fed by material parameters from high-fidelity, electronic structures level, density functional theory calculations, offering chemically accurate strength predictions with several orders lower computational costs. The physics of bifurcation in the evolution of brain morphologies is learned from simple sphere and ellipsoid models and then applied to predict the morphological development of human brains, showing excellent agreement with longitudinal magnetic resonance imaging (MRI) data. The learned latent variables are shown to be highly relevant to uncovered physical descriptors, explaining the effectiveness of the PT framework, which holds great potential in closing the gaps in understanding complexity problems in engineering sciences.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 14270