🔬 condensed matter

Machine Learning the H-theorem

This paper investigates the emergence of the thermodynamic arrow of time by training a permutation-invariant neural network on hard disk relaxation data to learn a scalar ordering of states, demonstrating that the model's learned function converges toward the structure of the Boltzmann H-functional without explicit physical constraints.

Original authors: Ruben Lier

Published 2026-07-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ruben Lier

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

The Mystery of the One-Way Street

Imagine you are watching a video of a glass shattering on the floor. It looks chaotic and messy. Now, imagine playing that video backward: the shards fly up, snap together, and land perfectly on the table. You know instantly that the backward version is fake. Why? Because in our universe, time has a direction. Things tend to go from order to disorder, like a tidy room slowly becoming a mess if you leave it alone. This is the heart of the Second Law of Thermodynamics, a rule that says the "messiness" (or entropy) of the universe always increases.

But here is the weird part: if you look at the tiny rules that govern individual atoms and molecules, they don't care about time. If you filmed two billiard balls bouncing off each other and played it backward, it would look exactly the same. The laws of physics are "time-reversible." So, how do we get a one-way street for time when every single step in the process is a two-way street? This puzzle has bothered scientists for over a century. They call it the "arrow of time." To solve it, they usually use a complex mathematical tool called the H-theorem, which acts like a scoreboard that always goes down as a system settles into equilibrium. But what if we didn't know the rules of the game? What if we just handed a computer a bunch of data and asked it to figure out which way time is flowing?

The Paper's Quest: Teaching a Robot to Feel Time

In this paper, the authors set out to see if a machine learning model could rediscover this "arrow of time" on its own, without being told the answer. They didn't give the computer the formula for the H-theorem or any physics textbooks. Instead, they created a virtual world filled with 1,000 hard disks (like tiny, perfect billiard balls) bouncing around in a square box. They started these disks in a chaotic, random state where they all moved at the same speed but in different directions. Then, they let the simulation run, recording the position and speed of every single disk at every moment.

The computer's job was simple but tricky: look at two snapshots of this bouncing gas—one from an earlier time and one from a later time—and guess which one came first. The only rule the computer was given was: "The later snapshot should get a higher score than the earlier one." To do this, the researchers built a special kind of neural network (a type of AI) called a "DeepSets" model. This network is designed to be "permutation-invariant," which is a fancy way of saying it doesn't care which particle is which. If you swap the names of two particles, the computer's answer stays the same, just like in real life where the physics doesn't change if you rename the atoms.

The model was trained using a "ranking loss," a scoring system that punished the computer if it guessed the wrong order. To make sure the computer didn't cheat by just giving every answer a tiny number, the researchers forced the scores to be standardized within each batch of data. They trained the model for 120 "epochs" (cycles through the data). At first, the computer was just guessing. But around the 60th epoch, something interesting happened: the computer suddenly got much better at its job. The error in its ranking dropped sharply, and it started to clearly distinguish between the "earlier" and "later" states.

The Big Reveal: Did It Find the H-Functional?

The most exciting part of the paper is what the computer actually learned. The researchers compared the "score" the computer gave to the snapshots with the actual H-functional, the famous mathematical formula Boltzmann invented in 1872 to measure how far a system is from equilibrium. They found that the computer's learned score matched the H-functional almost perfectly, once they adjusted for a simple shift in scale and position.

In other words, the AI, which was only told "later is bigger than earlier," independently figured out a mathematical function that behaves exactly like the H-theorem. It discovered that as the gas relaxes and the particles spread out their speeds into a smooth, predictable pattern (the Maxwell-Boltzmann distribution), the value of this function changes in a very specific, monotonic way. The paper shows that the dynamics of the system alone are enough to guide the model toward this structure.

However, the authors are careful to note that this result comes from a specific simulation of 1,000 hard disks with elastic collisions. They don't claim this proves the H-theorem for every possible system in the universe. They suggest that this approach works well for this specific setup, where the model learned to track the relaxation toward equilibrium without ever being shown the formula. They also found that the "temperature" parameter in their training loss mattered a lot: a lower temperature (0.1) made the model's output smooth and reliable, while a higher temperature (1.0) made it jittery and oscillatory near the end.

So, the paper doesn't solve the deep philosophical mystery of why time flows one way, but it does show that a machine, fed only with raw data and a simple ordering rule, can "learn" the same mathematical arrow of time that human physicists spent decades deriving. It suggests that the path to equilibrium is so strong and structured that even a neural network can find it just by watching the particles bounce.

Technical Summary: Machine Learning the H-theorem

Problem Statement
The H-theorem, formulated by Ludwig Boltzmann in 1872, provides a microscopic foundation for the Second Law of Thermodynamics by identifying a functional ( $H$ ) that decreases monotonically over time, thereby defining an arrow of time. However, a conceptual tension remains: the theorem relies on the molecular chaos assumption to derive irreversibility from microscopic interactions that are fundamentally time-reversal invariant. While recent machine learning approaches have explored irreversibility as a learning signal for sequential data (e.g., distinguishing forward/backward trajectories or learning entropy-like quantities), it remains an open question whether a generic neural network, trained solely on temporal ordering without prior knowledge of thermodynamics, can discover a scalar function that structurally resembles the H-functional.

Methodology
The authors investigate this question using a system of 1,000 identical hard disks in a periodic 2D box. The system is initialized with random positions and a uniform speed but random directions, creating a strongly non-equilibrium state that relaxes toward a Maxwell–Boltzmann distribution through deterministic, elastic collisions.

The machine learning architecture and training strategy are designed with three specific structural constraints:

Permutation Invariance: Since particle labels carry no physical meaning, the model utilizes the DeepSets architecture. Each particle's velocity is mapped to a hidden representation, which is then updated by combining the individual state with a global mean summary of the entire ensemble. This ensures the output is invariant to particle relabeling.
Weak Supervision via Ranking: The model is not provided with the analytical form of the H-functional. Instead, it is trained using a Siamese network structure where the same weights process configurations at time $t$ and $t+1$ . The objective is a ranking loss that forces the scalar output for the later state ( $t+1$ ) to be larger than the earlier state ( $t$ ).
Shrink-Free Loss: To prevent the model from trivially minimizing loss by shrinking the magnitude of all predictions, the authors employ a standardized softplus loss. Within each batch, outputs are standardized (mean and variance normalized) before calculating the ranking penalty. This ensures the model must learn a meaningful temporal progression rather than a scale collapse.

The model is trained for 120 epochs using the Adam optimizer on 500 statistically independent simulation runs, each containing 500 time steps.

Key Results

Emergence of a Monotone Scalar: The trained model successfully learns a scalar function that increases monotonically with time, effectively inferring an arrow of time directly from the microscopic velocity data.
Alignment with the H-Functional: When the learned scalar $h(V_t)$ is affinely fitted to the theoretical Boltzmann H-functional ( $H(t) = \int f \log f \, dv$ ), the two quantities show strong overlap. The model reproduces the overall relaxation trajectory of the system.
Sensitivity to Temperature Parameter: The study compares two models trained with different temperature parameters ( $\tau$ ) in the ranking loss. The model with $\tau = 0.1$ tracks the H-functional smoothly, particularly in the late-time regime near equilibrium. In contrast, a model with $\tau = 1$ exhibits visible oscillations around the equilibrium plateau, suggesting that the sharpness of the ranking penalty is quantitatively important for capturing the smooth relaxation dynamics.
Training Dynamics: The training history reveals a distinct transition around epoch 60, where the ranking loss decreases sharply and the mean normalized increment between time steps increases, indicating the model enters a regime of robust temporal ordering.

Significance and Claims
The paper claims to demonstrate that a permutation-invariant neural network, trained exclusively on the temporal ordering of microscopic states, can discover a scalar quantity that structurally mirrors the H-functional. This suggests that the arrow of time, as characterized by the H-theorem, is an emergent property that can be inferred directly from deterministic, reversible microscopic dynamics without explicit knowledge of the Boltzmann equation or the molecular chaos assumption.

The authors position this work as a step toward understanding how macroscopic irreversibility arises from microscopic reversibility through data-driven methods. They emphasize that their approach differs from prior work by handling unordered sets of particles directly and using a loss function that prevents trivial solutions. The study concludes that while the model successfully parametrizes the relaxation of this specific hard-disk system, future work is needed to determine if the learned quantity captures a general property of kinetic relaxation applicable to broader families of non-equilibrium initial conditions and more complex collision laws.

The Mystery of the One-Way Street

The Paper's Quest: Teaching a Robot to Feel Time

The Big Reveal: Did It Find the H-Functional?

Technical Summary: Machine Learning the H-theorem

More like this