+- MKLab (https://mklab.gr)
+-- Forum: [INDEX] (https://mklab.gr/forumdisplay.php?fid=1)
+--- Forum: ARTFICIAL INTELLIGENCE (AI) (https://mklab.gr/forumdisplay.php?fid=5)
+---- Forum: ARTICLES (https://mklab.gr/forumdisplay.php?fid=33)
+---- Thread: AI Method Tackles One of Science’s Hardest Math Problems (/showthread.php?tid=2193)
AI Method Tackles One of Science’s Hardest Math Problems - mklabgr - 09-25-2026
The University of Pennsylvania researchers have developed a new AI framework called Mollifier Layers for tackling inverse partial differential equations (PDEs), a particularly difficult class of mathematical problems. Ordinary PDEs describe how a system evolves from known rules—for example, how heat diffuses or how a population changes. An inverse PDE problem reverses this direction: given observations of a system, the goal is to infer the hidden parameters, forces, or processes that generated those observations. This makes inverse problems especially challenging because observations can be noisy and the underlying causes may be difficult to identify uniquely.
The central difficulty concerns differentiation. AI systems traditionally use recursive automatic differentiation to calculate derivatives through neural networks. For higher-order derivatives, however, this procedure can become unstable, particularly when the underlying data contain noise. Mathematically, if a neural network represents a function $u(x,t)$, a PDE might involve terms such as $\frac{\partial u}{\partial t}$, $\frac{\partial^2 u}{\partial x^2}$, or even higher-order derivatives. Repeatedly differentiating noisy approximations can amplify errors, making the inverse problem both computationally expensive and numerically unstable. The Penn team concluded that the bottleneck was not primarily the neural-network architecture but the mathematics of recursive differentiation itself.
Their solution draws on the mathematical concept of a mollifier, introduced by mathematician Kurt Otto Friedrichs in the 1940s. A mollifier smooths a function before its derivatives are calculated, suppressing sharp fluctuations and noise. Conceptually, instead of differentiating a noisy function $f(x)$ directly, one first constructs a smoothed function such as $f_\epsilon(x)$ and then computes $\frac{d}{dx}f_\epsilon(x)$. The Penn researchers incorporated this idea into neural networks as a mollifier layer. This allows the AI system to obtain higher-order derivatives more stably while substantially reducing the computational burden associated with recursive automatic differentiation.
One important application is the study of chromatin, the DNA-protein complex that packages genetic material inside cells. Researchers can observe chromatin structures but want to infer the hidden epigenetic processes and reaction rates that produce those structures. The new method potentially allows them to move from simply observing patterns to estimating the mathematical mechanisms responsible for them. The approach could also be relevant well beyond biology, including weather forecasting, materials science, fluid dynamics and other scientific fields in which researchers must infer hidden dynamics from noisy observations. The underlying research was published in Transactions on Machine Learning Research in March 2026 and is scheduled for presentation at NeurIPS 2026.
Key takeaways
- Inverse PDEs work backward: from observed effects to the hidden causes that generated them.
- A PDE may involve derivatives such as $\frac{\partial u}{\partial t}$ and $\frac{\partial^2u}{\partial x^2}$.
- Higher-order differentiation can amplify noise, creating numerical instability and high computational costs.
- Traditional AI approaches often rely on recursive automatic differentiation for these derivatives.
- Mollifiers smooth functions before differentiation, reducing the influence of noise.
- Penn's Mollifier Layers incorporate this classical mathematical idea directly into AI models.
- The important innovation is better mathematics rather than simply more computational power.
- A major biological application is inferring the processes controlling chromatin organization and epigenetic dynamics.
- The same framework could potentially be applied to weather, materials, fluids and other inverse scientific problems.
- The broader mathematical idea is powerful: smooth the data first, then differentiate, rather than allowing repeated differentiation to amplify noise.
ARTICLE