Engineers at the University of Pennsylvania have developed a novel AI method to solve inverse partial differential equations (PDEs), addressing a longstanding mathematical challenge with significant applications in genetics, weather forecasting, and materials science. The team, led by senior author Vivek Shenoy and including co-first authors Vinayak Vinayak and Ananyae Kumar Bhartari, introduced a technique called "Mollifier Layers," which was published in Transactions on Machine Learning Research and will be presented at the Conference on Neural Information Processing Systems in 2026. Inverse PDEs allow scientists to deduce hidden causes from observable effects, such as determining the source of a pebble dropped in a pond by analyzing the resulting ripples. While traditional differential equations model how systems evolve over time and space, inverse problems require working backward to infer the underlying dynamics. Researchers have historically struggled with these problems because standard AI approaches relying on recursive automatic differentiation become unstable and computationally expensive when dealing with noisy, higher-order data. This process often magnifies errors, similar to how repeatedly zooming in on a jagged line can distort the slope. To overcome this, the team adapted a mathematical concept called mollifiers, originally described in the 1940s. Mollifiers smooth out noisy or jagged functions by filtering sharp features before analysis. By inserting a "mollifier layer" into the neural network architecture, the researchers successfully smoothed the signal prior to differentiation. This innovation significantly reduced noise and lowered computational power requirements, allowing for more reliable solutions without needing to scale up hardware resources indefinitely. The immediate impact of this advancement is evident in the Shenoy Lab's work on chromatin, the bundled form of DNA inside cell nuclei. Understanding chromatin organization is crucial for comprehending gene expression, which governs cell identity, aging, and disease. Previously, researchers could observe chromatin structures but could not reliably infer the epigenetic reaction rates driving those formations. The new method enables the accurate modeling of how these chemical changes occur over time. By tracking how reaction rates evolve during aging, cancer, or development, scientists may eventually develop therapies that redirect cells to desired states by altering these rates. Beyond biology, the framework holds promise for other fields involving higher-order equations and noisy data, including fluid mechanics and materials science. The researchers aim to shift scientific inquiry from merely observing complex patterns to quantitatively uncovering the governing rules. As Shenoy noted, understanding the rules that generate a system provides the ability to change it, opening new avenues for discovery and intervention across diverse scientific domains. This approach marks a shift from relying solely on computational power to prioritizing improved mathematical methodologies in AI-driven science.