🤖 machine learning

Latent PDE mapping for efficient physics-informed learning across geometries with limited data

This paper introduces "latent PDE mapping," a physics-informed learning technique that transforms geometry-specific PDE residuals into a latent space to enable efficient generalization across diverse 2D and 3D geometries with sparse training data, achieving significant error reductions in solving the complex Aliev-Panfilov cardiac electrophysiology model.

Original authors: Ingvild Askim Adde, Mary M. Maleckar, Gabriel Balaban

Published 2026-07-27

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Ingvild Askim Adde, Mary M. Maleckar, Gabriel Balaban

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to predict how a ripple moves across a pond. If the pond is a perfect square, the robot learns the rules quickly. But what if the pond is a wobbly, stretching, twisting shape that changes every time you look at it? This is the daily struggle for scientists trying to use Artificial Intelligence (AI) to solve physics problems. These problems are governed by complex mathematical rules called Partial Differential Equations (PDEs), which act like the "laws of nature" for things like fluid flow, heat, or electricity. Usually, to train an AI to solve these laws for a specific shape, you need a massive amount of data. But in the real world—like designing a new airplane wing or simulating a patient's unique heart—getting that much data is impossible, expensive, or takes too long. Scientists have been stuck: either they use simple shapes where the AI works well, or they try to tackle complex, changing shapes and the AI gets confused, requiring them to retrain the whole system from scratch every time the shape changes.

This paper introduces a clever new trick called "Latent PDE Mapping" to solve this puzzle. Think of it as giving the AI a universal translator. Instead of trying to learn the physics of every weird, wobbly shape directly, the AI first translates the problem into a simple, standard "template" shape (like a perfect square or cube). It solves the physics there, where it's easy, and then translates the answer back to the weird shape. The authors tested this on the human heart, simulating how electrical waves travel through heart tissue—a task that is notoriously difficult because the waves move fast, change direction, and the heart itself is a complex, 3D organ that varies from person to person. They found that by using this "translation" method, the AI could learn from just fifteen different heart shapes and then accurately predict what would happen in completely new, unseen heart shapes. Without this trick, the AI struggled badly, especially when the heart was rotated or twisted. The study suggests that this method acts like a superpower for AI, allowing it to generalize its knowledge across different geometries with very little data, making it a promising tool for future engineering and medical simulations.

The Story of the Shape-Shifting Heart

Imagine you are a detective trying to solve a mystery: how does an electrical wave race through a heart? In the real world, every heart is different. Some are round, some are long, some are twisted like a pretzel. To simulate this on a computer, scientists use a set of strict rules (the Aliev-Panfilov equations) that describe how electricity moves through tissue. But here's the catch: if you change the shape of the heart even a little bit, the math changes completely.

Traditionally, if you wanted your AI detective to solve the mystery for a twisted heart, you'd have to feed it thousands of examples of twisted hearts. But what if you only have fifteen? That's the "limited data" problem. The AI usually gets confused, trying to memorize the specific shape instead of learning the actual rules of the game. It's like trying to learn how to drive by only practicing in a parking lot; when you hit a winding mountain road, you have no idea what to do.

The Magic Trick: The Universal Template

The authors of this paper came up with a brilliant workaround. They realized that instead of forcing the AI to learn the physics of every single weird shape, they could "pull back" the problem to a standard, pre-defined template shape. Let's call this template the "Latent Geometry."

Imagine you have a piece of clay in a weird, lumpy shape (the real heart). You want to know how electricity flows through it. Instead of calculating the flow on the lumpy clay directly, you imagine a magic map that flattens that lumpy clay into a perfect, smooth cube (the latent geometry). You solve the physics problem on the perfect cube, where the math is nice and easy. Then, you use a special "deformation gradient" (think of it as a stretchy rubber sheet) to map your answer back onto the lumpy clay.

The genius of this paper is that they didn't just map the shape; they mapped the laws of physics themselves. They showed that by translating the equations into this template world, the AI can learn the true rules of the game without getting distracted by the weirdness of the shape. This allows the AI to calculate "shape gradients"—a fancy way of saying it understands exactly how the solution changes when the shape changes. Without this, the AI was missing a huge chunk of information, like trying to drive a car while blindfolded to the turns.

The Heart Test: Twists, Turns, and Surprises

To test this idea, the researchers simulated the electrical activity of a heart using a computer model. They created two types of heart shapes:

The "Internal" Group: These were shapes the AI saw during training.
The "External" Group: These were brand new, unseen shapes the AI had never encountered, designed to test if the AI could truly generalize.

They tested four types of shape changes:

Expansion: The heart getting bigger or smaller.
Shearing: The heart getting squashed sideways.
Nonlinear: The heart getting wobbly and curved.
Rotation: The heart twisting around its axis.

The results were fascinating. When the heart just got bigger or squashed a little, the AI did okay even without the magic trick. But when the heart was rotated or twisted in complex ways, the difference was massive.

Without the "Latent PDE Mapping," the AI's predictions were a disaster. It produced solutions that looked nothing like the real physics, often showing the electrical wave getting stuck or moving in the wrong direction. It was like the AI was guessing wildly.

With the "Latent PDE Mapping," the AI became a pro. On the most difficult rotated shapes, the new method reduced the error by a factor of 4 to 6. In plain English, the AI's predictions became four to six times more accurate. The electrical waves it predicted looked almost identical to the "ground truth" (the perfect computer simulation), preserving the correct pattern of the wave moving across the heart.

Why This Matters (And What It Doesn't)

The authors are careful to point out that this isn't a magic wand that fixes everything instantly. The method works best when the shape changes cause big, dramatic shifts in how the physics behaves—like when a rotation changes the direction of the heart's muscle fibers. If the shape change is small and simple, the AI can sometimes get away with the old methods.

Also, there is a small cost. To set up this "magic map," the computer has to do some extra math before the training even starts. It takes a few seconds to calculate the stretchy map for each shape. However, once the AI is trained, using it to make predictions is just as fast as the old methods. The extra work is a one-time setup fee, not a slow-down for every prediction.

The study suggests that this approach is a powerful new tool for "Physics-Informed Machine Learning." It allows scientists to build models that are robust and generalizable, even when they only have a tiny amount of data. While the paper focused on heart electricity, the authors hint that this "template" idea could work for other complex problems, like designing airplane wings or understanding how blood flows through different arteries.

In the end, this paper shows that by teaching AI to see the world through a standardized lens, we can help it understand the messy, changing reality of our physical world with far fewer examples than we thought possible. It's a step toward making AI that doesn't just memorize the data, but truly understands the rules of the game, no matter how the board is shaped.

Technical Summary: Latent PDE Mapping for Efficient Physics-Informed Learning Across Geometries

Problem Statement
Physics-informed machine learning (PIML) offers a powerful paradigm for solving complex physical problems by embedding governing equations as soft constraints. However, a significant limitation arises when applying these models to diverse or novel geometries. Conventional PIML formulations require costly retraining for every new geometry, undermining their computational efficiency. While specialized architectures (e.g., graph neural networks) or input augmentations (e.g., shape descriptors) have been proposed to handle variable geometries, they often struggle with generalization, require complex meshing, or fail to accurately capture the geometric dependency within the physics loss gradients. Specifically, standard formulations often neglect the contribution of boundary motion to the shape gradient, leading to inaccurate gradient estimates during training and poor generalization to unseen shapes, particularly in nonlinear, time-dependent systems with sharp gradients.

Methodology
The authors introduce Latent PDE Mapping (LPM), a framework designed to enable efficient geometric generalization with sparse training data. The core concept involves pulling back geometry-specific Partial Differential Equation (PDE) residuals and boundary conditions (BCs) from a physical geometry $\Omega(s)$ to a predefined, shared latent geometry $\Omega_0$ via a diffeomorphic mapping $\Phi$ .

Mathematical Formulation: The method leverages principles from nonlinear solid mechanics and the Finite Element Method (FEM). It utilizes the deformation gradient $F$ and the deformation Jacobian $J$ to map the PDE residuals from the physical domain to the latent domain. For the Aliev-Panfilov model of cardiac electrophysiology, the governing equations are transformed such that the dependence on shape parameters $s$ is moved from the domain definition into the PDE operators themselves (e.g., transforming the diffusion term $\nabla \cdot (D\nabla V)$ into $\frac{1}{J}\nabla \cdot (JF^{-1}DF^{-T}\nabla V)$ ).
Gradient Accuracy: By mapping the problem to a fixed latent geometry, the integrand of the physics loss becomes independent of the domain boundary movement. This allows for the direct calculation of shape gradients $\partial L_{phys} / \partial s$ without neglecting the boundary motion term (the second term in the Leibniz integral rule), which is typically omitted in discrete loss formulations on varying geometries.
Implementation: The framework is architecture-agnostic but was implemented and tested using two specific models:
1. Physics-Informed Neural Networks (PINNs): Fully connected networks with shape parameters or PCA descriptors as inputs.
2. Physics-Informed Deep Operator Networks (PI-DONs): Branch-trunk architectures mapping initial conditions and geometry descriptors to solution operators.
Experimental Setup: The study utilized the anisotropic, nonlinear, time-dependent Aliev-Panfilov PDE as a benchmark. Training data was generated via FEM simulations (openCARP) on parameterized 2D and 3D geometries (expansion, shear, nonlinear deformation, and rotation). The models were trained on limited data (15 geometric samples) and tested on both internal (training distribution) and external (out-of-distribution) geometric families.

Key Contributions

Novel Mapping Technique: Introduction of Latent PDE Mapping, which maps geometry-specific PDE and BC residuals to a shared latent geometry, enabling physics-informed frameworks to learn meaningful representations across diverse shapes using minimal data.
Theoretical and Empirical Validation: Demonstration that LPM correctly accounts for geometric variability in the physics loss gradient. The authors provide theoretical evidence that this leads to more accurate shape-gradient calculations and empirical evidence showing improved generalization.
Broad Applicability: The framework is shown to be effective across different neural network backbones (PINNs and PI-DONs) and in both 2D and 3D settings, handling complex, nonlinear, and time-dependent PDEs.

Results

Performance Gains: In 2D experiments, LPM-enhanced models (LPM-PINN and LPM-DON) consistently outperformed ablation models (which used latent geometries without the PDE pull-back or standard PINNs) across 11 out of 16 and 14 out of 16 experiments, respectively.
Error Reduction: The most significant improvements were observed in rotational and nonlinear deformation families, where LPM achieved a ~4-6x reduction in mean relative $L_2$ error compared to the second-best performing models. For example, on the external rotation family ( $G^*_{rot}$ ), LPM-PINN reduced the error from 0.289 (LG-PINN) to 0.068.
3D Generalization: In 3D rotational experiments, LPM models significantly outperformed ablation models, particularly for rotations about the y- and z-axes ( $H^*_{y,rot}, H^*_{z,rot}$ ), where fiber orientation changes induced substantial solution variations.
Data Efficiency: The models achieved these results using only 15 training geometries, demonstrating efficacy in the limited data regime.
Computational Cost: The additional computational cost of LPM was modest, confined primarily to the training phase (precomputing deformation gradients). Inference times remained comparable to ablation models.

Significance and Claims
The paper claims that Latent PDE Mapping provides a mathematically rigorous and broadly applicable framework for extending physics-informed learning to diffeomorphic varying geometries. By transforming a geometry-dependent learning problem into a parametric one on a fixed latent domain, the method facilitates the automated calculation of geometry-consistent shape gradients that are missing in conventional formulations.

The authors emphasize that the technique is particularly valuable for scenarios where geometric variations induce substantial changes in PDE solutions (e.g., rotations altering anisotropy directions). While the method shows promise for complex, nonlinear, and dynamic systems, the authors remain modest, noting that the approach currently assumes knowledge of the exact deformation gradient and has been validated primarily on a single PDE (Aliev-Panfilov). They suggest that future work is needed to address the approximation of deformation gradients from raw geometric data and to validate the method on broader classes of PDEs and time-dependent deformations. Ultimately, the study highlights that LPM enables the creation of generalizable physics-informed models from limited training sets, addressing a critical bottleneck in applying PIML to real-world engineering and medical problems involving diverse geometries.

The Story of the Shape-Shifting Heart

The Magic Trick: The Universal Template

The Heart Test: Twists, Turns, and Surprises

Why This Matters (And What It Doesn't)

More like this