Physics-Augmented Learning: A New Paradigm Beyond Physics-Informed Learning

The Big Picture

Imagine teaching someone to paint portraits. One approach: show them a finished painting, then correct every brushstroke that violates proportion. Another: hand them a brush that physically cannot make out-of-proportion marks.

Both methods enforce the same rules, but they work differently — and for different kinds of rules, one will outperform the other. This tension runs through how physicists and machine learning researchers have been trying to teach neural networks about the laws of nature.

The dominant strategy has been physics-informed learning (PIL), which embeds physical knowledge into a model’s training by penalizing outputs that violate known physics. Enforce energy conservation by punishing solutions that break it. Enforce governing equations by adding a penalty whenever the model strays from them. Powerful, widely used, and well-understood.

But PIL carries a hidden assumption: it only works when you can efficiently check whether a solution violates the physics. Not all physical properties are checkable this way.

Ziming Liu, Max Tegmark, and collaborators at MIT and IAIFI identified this blind spot and built a complementary framework to fix it. Their proposal, physics-augmented learning (PAL), flips the script: instead of penalizing violations, PAL embeds physical properties directly into the model’s architecture, making violations structurally impossible from the start.

Key Insight: Physics-informed learning can only enforce properties you can test for — but some of the most important physical structures, like Lagrangian dynamics and hidden symmetries, can be generated but not efficiently tested. PAL handles exactly these cases.

How It Works

The paper draws a conceptual line that has real practical consequences. Physical properties fall into two categories:

• Discriminative properties — those for which you can write an efficient test that detects violations. If a function satisfies the property, the test returns zero; if not, it returns something nonzero.
• Generative properties — those for which you can construct objects satisfying the property, but have no efficient way to check whether an arbitrary object satisfies it.

This distinction maps directly onto how machine learning can use each type. PIL exploits discriminative properties: it adds a penalty term to the training loss using a discriminator operator, essentially a mathematical test for whether the physics is satisfied. For differential equation constraints, separability conditions, or directly visible symmetries, this works well.

Now consider the Lagrangian property, whether a given set of equations of motion can be derived from an underlying energy function (a Lagrangian). Generation is trivial: pick a Lagrangian and derive the dynamics. But determining whether an arbitrary acceleration field is Lagrangian has no known efficient solution. PIL simply cannot enforce this structure.

The same goes for positive definiteness, the guarantee that a quantity like energy always stays non-negative. And for hidden symmetries, patterns in the physics that aren’t obvious from the surface form of equations but deeply constrain the solutions. In all these cases, you can build objects with the property, but you can’t cheaply test whether a given object has it.

PAL’s solution is architectural. Rather than starting with a generic network and constraining it, PAL builds the property into the network’s structure. For Lagrangian systems, the model directly parameterizes the Lagrangian function $L(q, \dot{q}; \theta)$ and derives accelerations via the Euler-Lagrange equations, the classical rules connecting an energy function to the motion it implies. The model cannot output non-Lagrangian dynamics, not because it’s penalized, but because the architecture makes it impossible.

The paper uses additive separability, the property that $f(x_1, x_2) = f_1(x_1) + f_2(x_2)$, as a benchmark where both approaches go head-to-head. Separability is both discriminative (via the cross-derivative test $\partial^2 f / \partial x_1 \partial x_2 = 0$) and generative, so it allows a direct comparison. In the PAL approach, the network is structured as $f_1(x_1; \theta_1) + f_2(x_2; \theta_2) + f_{12}(x_1, x_2; \theta_{12})$, with a loss penalizing the residual $f_{12}$ term. PIL uses a single network with a cross-derivative penalty. Both work here. For purely generative properties, only PAL succeeds.

The full taxonomy covers Lagrangian structure, positive definiteness, manifest symmetry, hidden symmetry, separability, and PDE satisfiability, organized into a table showing which properties are generative, discriminative, or both. PIL covers the discriminative column. PAL covers the generative column. Together, they span the full space.

Why It Matters

Much of modern physics involves quantities that are easy to construct but hard to verify. Action principles say physical systems follow paths that minimize a certain quantity. Gauge symmetries encode real physical structure through mathematical redundancies in how we describe forces. Symplectic structures and integrability conditions impose powerful constraints on dynamics. PIL’s penalty-based approach doesn’t apply to any of them. PAL provides the missing piece.

There’s a practical efficiency argument too. Even when a property is discriminative and PIL could in principle be used, a PAL-style architectural embedding may train faster and generalize better, because the constraint is exact by construction rather than approximately enforced through a tuned penalty coefficient. The paper shows cases where PIL fails to converge or requires enormous penalty weights, while PAL succeeds out of the box.

Neural networks are pushing deeper into scientific computing: modeling molecular dynamics, predicting gravitational wave signals, emulating cosmological simulations. Whether these models are trustworthy depends increasingly on their ability to encode rich physical structure faithfully.

The framework also points toward a clear research agenda. One direction is systematically cataloguing physical properties by their generative and discriminative nature. Another is developing PAL architectures for increasingly complex structures, or hybrid approaches combining both paradigms. The connections to equivariant neural networks, Hamiltonian neural networks, and neural ODEs are natural starting points, since all three already embed specific physical constraints into network design.

Bottom Line: PAL fills a gap PIL cannot. By building physical properties into network architecture rather than penalizing their violation, it handles the broad class of generative physical structures like Lagrangian dynamics and hidden symmetries that conventional physics-informed methods simply can’t reach.

---