A pilot study assessing the feasibility of integrating PINNs in the biomedical curriculum

Abstract

To address the growing need for interdisciplinary skills in biomedical engineering education, this paper introduces a longitudinal hands-on learning experience centering on Physics-Informed Neural Networks. We developed a two-stage pedagogical pathway following the same cohort of students from their second to their third year, designed to demystify the convergence of machine learning and computational mechanics.

Initially, within a second-year continuum biomechanics course, students established the fundamentals of the methodology by solving the Fisher equation and a damped harmonic oscillator. This phase allowed learners to transition from conceptual understanding to the basic implementation of physical laws within deep learning frameworks.

Subsequently, in the third year, an advanced activity focused on a mechanical problem of linear elasticity was introduced. In this phase, students were challenged to autonomously generate synthetic ”ground truth” data through Finite Element Method simulations using the open-source software FEBio. This activity served a dual pedagogical purpose: it reinforced the application of PINNs in solid mechanics and, simultaneously, empowered students to create and validate against finite element models.

Quantitative analysis of pre- and post-intervention tests, evaluated via the Wilcoxon signed-rank test, shows a large effect size in knowledge acquisition regarding PINN fundamentals. Furthermore, technical assessment of the hybrid projects using a specific rubric reveals that students successfully navigated the interoperability between FEM and Deep Learning. We discuss the potential of this ”Gray-Box” educational model to modernize biomechanics instruction and its broader implications for other engineering disciplines, while acknowledging the limitations of a single-cohort design regarding causal inference and generalizability.