Engineers at the University of Pennsylvania have introduced a groundbreaking method known as “Mollifier Layers” to address the complexities of inverse partial differential equations (inverse PDEs). These equations are essential across various scientific domains including weather systems, biology, and materials science, yet they present significant challenges as they require deducing underlying causes from observable phenomena.
The researchers plan to showcase their findings at NeurIPS 2026, marking a shift from conventional AI methods that emphasize model size and computational strength. Senior author Vivek Shenoy, who holds the title of Eduardo D. Glandt President’s Distinguished Professor in Materials Science and Engineering, likens the challenge of inverse PDEs to the difficulty of identifying the source of ripples in a pond.
Traditional approaches in AI have struggled with the differentiation process, which is crucial for measuring changes in these complex systems. The Penn team noted a notable increase in memory usage—from 0.21 gigabytes to 2.70 gigabytes—when analyzing a fourth-order reaction-diffusion problem. By integrating mollifiers, a mathematical tool that smooths functions, they aim to enhance the stability of derivative calculations and improve the efficiency of their model.