Engineers at the University of Pennsylvania have created a new artificial intelligence (AI) method to solve inverse partial differential equations. These equations form a hard class of mathematical problems. Inverse partial differential equations, or inverse PDEs, let scientists start with visible patterns and figure out the unseen forces or rules that caused them. The new technique is called mollifier layers and could help many areas including genetics and weather forecasting.
Partial differential equations, or PDEs, describe how things change across space and time, such as heat moving through a material or weather systems developing. Inverse PDEs reverse the process. They work from the final result back to the original causes. An example is seeing ripples on a pond and determining exactly where and when a pebble landed.
Improved mathematics for cleaner and more efficient results
The researchers found that common AI methods become unstable with complex or noisy information because they repeatedly calculate changes in a process that can magnify small errors. To fix it, they added mollifier layers. Mollifiers are mathematical tools developed in the 1940s that gently smooth out rough or noisy data before measuring changes. This makes the calculations more stable and uses less computing power.
One important use is in the study of chromatin. Chromatin is the mixture of DNA and proteins that packages genetic material inside cell nuclei. The new method helps scientists infer the hidden chemical reaction rates that control how chromatin folds and which genes are active. These small structures, only 100 nanometers wide, influence cell behavior, aging, and disease. Better understanding them could lead to new medical treatments.
The approach may also improve work in materials science, fluid mechanics, and other fields that deal with complex systems and imperfect data. By moving from simply observing patterns to uncovering the underlying rules, the technique offers a path to better prediction and control of natural processes. The study will be published in Transactions on Machine Learning Research.