Sometimes scientific progress requires better mathematics, not more computing power.
At the University of Pennsylvania, a team of engineers has revived a mathematical idea from the 1940s to help artificial intelligence do what scientists have always longed to do: reason backward from visible effects to hidden causes. Their technique, called Mollifier Layers, tames the noise that has long destabilized AI attempts to solve inverse partial differential equations—problems that appear wherever nature conceals its mechanisms behind its outcomes. The work is a reminder that progress sometimes demands not more computational force, but a more honest reckoning with the mathematics underneath.
- Inverse PDEs have resisted AI for years because standard methods amplify noise with each calculation step, making complex systems increasingly unstable and expensive to model.
- The Penn team traced the failure not to their neural network's design but to the mathematical foundation itself—recursive automatic differentiation—a diagnosis that reframed the entire problem.
- By embedding Kurt Otto Friedrichs's 1940s smoothing tools directly into the AI architecture, they intercepted noise before it could compound, slashing both error and computational cost in a single move.
- The immediate target is chromatin—the nanoscale DNA packaging that governs gene activity, cell identity, aging, and disease—where the technique can now infer hidden chemical reaction rates rather than merely photograph static structure.
- Researchers believe mollifier layers could propagate across materials science, fluid mechanics, and weather forecasting, wherever noisy data obscures the rules driving a system.
A team at Penn has fused a forgotten mathematical tool with modern AI to crack a category of problems that has long frustrated science: inverse partial differential equations, which ask researchers to reason backward from observable effects to the hidden causes behind them. As team leader Vivek Shenoy puts it, the challenge is like studying ripples in a pond and working backward to find where the pebble fell.
For years, AI systems approached these problems through recursive automatic differentiation—repeatedly computing how quantities shift through a neural network. In noisy, high-order systems, however, each additional step amplifies imperfection rather than resolving it, and the computational cost grows accordingly. The Penn researchers eventually realized the problem was not the network's architecture but the mathematical method itself.
Their solution came from an unlikely era. In the 1940s, mathematician Kurt Otto Friedrichs developed "mollifiers"—tools that smooth jagged or noisy functions by softening their sharpest features. By inserting a mollifier layer into their AI system to smooth signals before measuring change, the team dramatically reduced both noise and computational burden.
The most immediate application is chromatin, the nanoscale packaging of DNA and proteins that controls which genes are active inside a cell. Understanding how its chemical reaction rates evolve during aging, cancer, or development could open pathways to therapies that redirect cells toward healthier states—moving science from passive observation to active intervention.
Beyond genetics, the researchers see mollifier layers as a broadly transferable template for any field—materials science, fluid mechanics, meteorology—where hidden parameters must be inferred from noisy data. Their deeper argument is philosophical as much as technical: sometimes the path forward requires better mathematics, not more computing power. "If you understand the rules that govern a system," Shenoy says, "you now have the possibility of changing it."
A team of engineers at Penn has borrowed a mathematical technique from the 1940s and woven it into modern artificial intelligence to solve a class of problems that has long resisted easy answers. The breakthrough, which they call "Mollifier Layers," addresses inverse partial differential equations—a category of mathematical challenges that asks scientists to work backward from what they can see to figure out what caused it. The work will be presented this year at a major machine learning conference.
The practical stakes are high. Inverse problems appear everywhere in science. A geneticist looking at how DNA folds inside a cell nucleus wants to know what chemical forces are at work. A meteorologist studying weather patterns wants to infer the hidden dynamics driving them. A materials scientist examining how heat moves through a substance wants to reverse-engineer the underlying physics. In each case, the observable outcome is clear—the hidden mechanism is not. "Solving an inverse problem is like looking at ripples in a pond and working backward to figure out where the pebble fell," explains Vivek Shenoy, a materials science professor who led the research. "You can see the effects clearly, but the real challenge is inferring the hidden cause."
For years, artificial intelligence systems have tackled these problems by computing derivatives—mathematical measures of how things change—through a method called recursive automatic differentiation. The approach works by repeatedly calculating how quantities shift as they move through a neural network. But when dealing with complex, higher-order systems contaminated by noisy data, this method becomes unstable. Each additional calculation step can actually amplify the noise rather than reduce it, like zooming in repeatedly on a jagged line until the imperfections overwhelm the signal. The computational cost balloons accordingly. The Penn team realized the problem was not a flaw in the neural network's design but in the mathematical foundation itself.
The solution came from an unexpected direction. In the 1940s, mathematician Kurt Otto Friedrichs developed what he called "mollifiers"—mathematical tools designed to smooth out jagged or noisy functions by softening their sharpest features. The Penn researchers adapted this decades-old technique for modern machine learning. By inserting a "mollifier layer" into their AI system, they could smooth the signal before measuring how it changes. This single modification dramatically reduced both the noise in the results and the computational power required to achieve them. "We initially assumed the issue had to do with the neural network's architecture," says Ananyae Kumar Bhartari, one of the paper's co-authors. "But after carefully adjusting the network, we eventually realized the bottleneck was recursive automatic differentiation itself."
The immediate application driving this research is chromatin—the intricate packaging of DNA and proteins that stores genetic material inside cells. These structures exist at a scale of just 100 nanometers, yet they control which genes are accessible and therefore active. That accessibility determines cell identity, function, aging, and disease. Shenoy's lab has long studied how epigenetic processes—chemical changes that regulate gene activity without altering DNA itself—organize chromatin structure. With mollifier layers, researchers can now infer the rates at which these chemical reactions occur, moving beyond simply observing chromatin's structure under a microscope to modeling how it evolves over time. "If we can track how these reaction rates evolve during aging, cancer, or development," says Vinayak Vinayak, a doctoral candidate on the team, "this creates the potential for new therapies: If reaction rates control chromatin organization and cell fate, then altering those rates could redirect cells to desired states."
The implications extend far beyond genetics. Inverse problems pervade materials science, fluid mechanics, and numerous other fields where scientists must infer hidden parameters from observable data. Any domain dealing with higher-order equations and noisy measurements could benefit from a more stable, computationally efficient mathematical framework. The researchers see mollifier layers as a template—a demonstration that sometimes scientific progress requires not more computing power but better mathematics. "Ultimately, the goal is to move from observing complex patterns to quantitatively uncovering the rules that generate them," Shenoy says. "If you understand the rules that govern a system, you now have the possibility of changing it."
Notable Quotes
Solving an inverse problem is like looking at ripples in a pond and working backward to figure out where the pebble fell.— Vivek Shenoy, Penn Materials Science professor
Some scientific challenges require better mathematics, not just more compute.— Vinayak Vinayak, doctoral candidate in Materials Science and Engineering
The Hearth Conversation Another angle on the story
Why does this matter? Inverse problems sound abstract.
They're not abstract at all. Every time a doctor wants to understand what's causing a disease, or a climate scientist wants to know what forces are driving a weather pattern, they're solving an inverse problem. They see the outcome; they need to infer the cause.
So the old AI methods couldn't do this reliably?
They could, but at a cost. Recursive automatic differentiation works by repeatedly measuring change through a neural network. With noisy data, each step amplifies the noise. It's like trying to measure the slope of a jagged line by zooming in closer and closer—eventually the imperfections overwhelm the signal.
And mollifier layers fix that how?
By smoothing the signal first, before you measure it. It's a 1940s mathematical technique that was sitting there, unused in this context. Once they adapted it, the noise dropped dramatically and so did the computational burden.
What's the real-world payoff?
In chromatin research, it means they can now figure out how fast the chemical reactions are happening that control which genes turn on or off. That's the missing piece. If you know the reaction rates, you might be able to change them—to redirect cells toward a healthier state.
Is this just for biology?
No. Materials science, fluid mechanics, weather modeling—anywhere you have noisy data and complex systems. The mathematical framework is general. That's what makes it powerful.
So this is a tools paper, not a discovery paper?
It's both. They've created a tool, yes. But the discovery is that sometimes the bottleneck isn't the AI architecture—it's the math underneath it. That changes how people think about these problems.