Artificial intelligence is often described as transforming science through speed and automation, yet its real impact has been more measured. The reason is not a lack of expressive power, but a structural mismatch between how machine learning systems learn and how scientific knowledge is organized.

Scientific theories succeed because they encode general constraints which apply broadly—across elements, environments, sizes and time scales. Physical and chemical laws define what is allowed, not just what has been observed, enabling extrapolation. Most machine learning models, by contrast, learn correlations without an inherent notion of physical constraints, making extrapolation their central failure mode.

At the same time, many scientific systems are fundamentally high dimensional and computationally intractable using traditional methods. Molecular dynamics provides a clear example. Every initial configuration defines a trajectory through an astronomically large configuration space, yet no successful scientific method attempts to represent this space explicitly.

What matters instead is the global structure of that space. From a thermodynamic perspective, the object of interest is not a single trajectory but an ensemble. Boltzmann statistics describes how probability mass is distributed across configurations, defining stable basins, free energy barriers, and long time behavior.

This reframes what it means to design machine learning for molecular dynamics. The goal is not to reproduce exact atomic trajectories, which merely recreates the original computational bottleneck. The goal is to learn a representation whose geometry reproduces the correct ensemble while compressing irrelevant microscopic detail.

Under this framing, machine learning is useful not because it replaces physics, but because it can learn compact coordinates for physically admissible spaces that are too complex to explore directly. Physics defines the constraints, while ML learns how real systems occupy the constrained landscape. Extrapolation becomes meaningful only within this admissible space.

Several recent works explicitly adopt this perspective. Energy-based coarse-graining approaches that target the Boltzmann distribution treat the equilibrium ensemble as the object being learned, rather than individual trajectories. These models define learned energy landscapes whose induced probability measures match atomistic ensembles after marginalization.

Critically, these approaches succeed precisely because they constrain learning to a physically admissible hypothesis space. At the same time, their limitations are instructive. Correctness is guaranteed only with respect to the chosen coarse representation, dynamics are not uniquely determined by the equilibrium ensemble, and extrapolation remains bounded by assumptions about transferability of the coarse variables. These works therefore illustrate the core principle rather than resolve it: machine learning becomes scientifically meaningful when it learns representations of admissible ensembles, but its success depends entirely on how those representations are defined.

This example generalizes beyond molecular dynamics. For any scientific domain, the design question is the same. What object does the science actually care about, and what constraints define admissible behavior? Machine learning becomes powerful when it learns efficient representations of those constrained spaces, not when it attempts to infer the world from data alone.