How AI Just Made the Impossible Possible in Molecular Science

99.8% of simulated molecular transitions now succeed—up from 0%—using a new physics-informed AI method.

That’s not a typo. In one benchmark, every single attempt to simulate a molecule jumping between stable states failed under normal conditions. Not one in a thousand made it across. But with a new technique that fuses dynamical systems theory with optimal control, 998 out of 1,000 trajectories crossed successfully within the same number of steps. This isn’t incremental progress. It’s a leap from invisibility to clarity.

The problem lies in how molecules move. Proteins fold, drugs bind, enzymes activate—all through tiny structural shifts that unfold over time. But these changes often require crossing high energy barriers, like rolling a ball over a mountain rather than around it. Under natural dynamics, such transitions are so rare they might take years to observe in simulation. Yet understanding them is essential for designing better medicines, materials, and catalysts.

For decades, scientists have relied on enhanced sampling methods to speed things up. Some add artificial forces. Others use machine learning to learn bias potentials from data. But most require either hand-tuned parameters or, worse, initial examples of the very transitions they’re trying to find—examples that may not exist.

Now, Boya Hou, Shane Wang, and their colleagues at the University of Illinois Urbana-Champaign have introduced a fundamentally different approach: Transition Path Sampling Using Koopman Operators and Exit-Time Optimal Control (Hou et al., 2026). Their method doesn’t guess. It computes.

The Science

At the heart of this breakthrough is a shift in perspective—from simulating fixed-time trajectories to controlling systems until they hit a target.

Traditional methods treat transition path sampling (TPS) as a problem over a fixed time window. You run a simulation for, say, 1 nanosecond, and hope something interesting happened. But real transitions don’t obey clocks. They’re random events whose timing depends on the path itself. Fixing the horizon creates tension: too short, and nothing happens; too long, and you waste computation.

Instead, Hou et al. frame TPS as an optimal stochastic control (OSC) problem up to an exit time—specifically, the first time a trajectory reaches the target state. This makes the time horizon itself a random variable, aligned with the physics of rare events.

They begin with the Gibbs variational principle, which links free energy minimization to probability measures on path space. By defining an energy functional that penalizes time spent far from the transition pathway, they identify a target path measure—the one that maximizes the likelihood of observing reactive trajectories while staying close to the original dynamics.

But identifying the right measure isn’t enough. You still need to sample from it.

Here’s where their innovation shines: they prove that the optimal controller—the force you must apply to guide the system efficiently to the target—has a closed-form solution:

$$u^{\ast }(x)=\nabla \log \rho (x)$$

where $\rho(x)$ is the expected value of the exponential of the negative cumulative cost along uncontrolled paths starting at $x$. Crucially, $\rho(x)$ solves a linear boundary value problem governed by the Koopman generator, a differential operator tied to the underlying dynamics.

This means the entire control policy can be computed without running a single biased simulation.

To make this practical, they approximate $\rho(x)$ in a reproducing kernel Hilbert space (RKHS), turning the problem into a single equality-constrained quadratic program. No gradient descent. No backpropagation. No simulation loops during training. Just one convex optimization problem whose solution gives the full control field.

And because the Koopman operator is linear—even when the dynamics aren’t—its spectrum reveals the slowest modes of the system. The leading eigenfunctions expose metastable sets (like folded and unfolded protein states) and allow estimation of the committor function $q(x)$, which gives the probability that a trajectory starting at $x$ will reach the target before returning to the source.

Remarkably, this can be done using only data from non-reactive regions. No transition paths required.

What They Found

The authors tested their method on two canonical systems:

1. Two-channel double well: a 2D potential with two minima connected by two narrow channels (

). Standard simulations starting in one basin never cross to the other within 1,000 steps. 2. Alanine dipeptide: a small peptide used as a benchmark in molecular dynamics, with conformational states defined by its backbone dihedral angles $\phi, \psi$ (the Ramachandran plot).

In both cases, uncontrolled dynamics yielded 0% success rate—not a single trajectory reached the target within the allotted time.

With their method, success rates jumped to:

• 99.8% in the double well within 1,000 steps
• 93% in alanine dipeptide within 1 picosecond

Success Rate of Transition Path Sampling Methods

Label	Value
Uncontrolled	0
TPS-DPS	0
Proposed Method	0

Source: Hou et al., 2026. Uncontrolled dynamics fail completely; TPS-DPS reaches 42%; the proposed method achieves 99.8%.

shows how dramatically this outperforms both uncontrolled dynamics and TPS-DPS, a recent neural-network-based method that achieves only 42% success in the double well.

Cumulative Transition Probability Over Time

Label	Value
0.1 ps	0
0.3 ps	0
0.5 ps	0
0.7 ps	0
1.0 ps	0

Source: Hou et al., 2026. The proposed method achieves 93% transition probability within 1 picosecond, compared to 0% for uncontrolled dynamics.

illustrates the rapid rise in transition probability over time for alanine dipeptide—while uncontrolled dynamics remain flat at zero, the controlled system quickly climbs toward near-certainty.

visualizes the key ingredients: the first nontrivial Koopman eigenfunction cleanly separates the C5 and C7$_{\mathrm{ax}}$ basins, and the estimated committor smoothly interpolates between them, guiding the controller.

Critically, the method doesn’t just produce more transitions—it produces physically meaningful ones. Because the controlled dynamics are absolutely continuous with respect to the original process, the authors show (via Lemma 4) that trajectories can be reweighted to recover the statistics of the true, unbiased dynamics. This means you can compute correct free energy differences, transition rates, and mechanistic insights—not just fast cartoons.

Why This Changes Things

Rare event sampling sits at the core of some of science’s hardest problems. Protein folding. Chemical reaction rates. Nucleation in phase transitions. All depend on processes that happen once in a blue moon—but whose consequences shape biology, industry, and climate.

Current deep learning approaches, while powerful, suffer from several limitations:

• They require many training iterations with biased rollouts, which may never encounter the target.
• They lack theoretical guarantees about convergence or optimality.
• They’re often black boxes, making it hard to interpret why a path was taken.

Hou et al.’s method bypasses these issues entirely. By grounding the solution in the Koopman generator—a well-studied object in dynamical systems theory—they build a bridge between modern machine learning and classical physics.

The fact that the optimal controller emerges from a linear PDE is profound. Nonlinear control problems are notoriously hard. But here, linearity arises not from simplifying assumptions, but from the structure of the problem itself: the committor satisfies $L q = 0$, and $\rho$ satisfies $L\rho - f\rho = 0$, both linear equations involving the infinitesimal generator $L$.

This opens the door to interpretability. The control field isn’t a tangle of neural net weights—it’s literally the gradient of a physically meaningful quantity: the log-expectation of path cost. Where the committor is low (deep in the source basin), the penalty is high, so the controller pushes hard. Near the target, the push fades, avoiding overshoot.

Moreover, because the method operates in RKHS, it inherits the benefits of kernel methods: flexibility, robustness, and strong theoretical foundations. And unlike neural networks, there’s no risk of poor initialization or vanishing gradients.

Perhaps most importantly, this work challenges the assumption that you need examples to learn rare events. Traditional wisdom says you can’t train a model to recognize what you’ve never seen. But Koopman spectral analysis extracts global dynamical structure from local data—even if no trajectory ever crosses the barrier. It’s like deducing the existence of a bridge by studying the banks of a river.

What's Next

This framework is promising—but not yet a panacea.

First, scalability remains a question. While solving a quadratic program is much cheaper than training a neural network, RKHS methods typically scale cubically with the number of basis functions or data points. For large molecular systems (thousands of atoms), direct application may be prohibitive unless combined with dimensionality reduction or sparse approximations.

Second, the current formulation assumes known drift and diffusion coefficients ($b(x)$ and $\sigma(x)$). In complex systems like proteins, these may not be available analytically. However, recent advances in learning SDEs from data (e.g., via score matching or diffusion models) could fill this gap.

Third, while the method excels at two-state transitions, real biomolecular processes often involve multiple intermediates. Extending this approach to multi-basin landscapes—perhaps using higher-order Koopman eigenfunctions to define a hierarchy of committors—is a natural next step.

Finally, experimental validation is crucial. Simulations can be tuned to perfection, but biology is messy. Testing whether these predicted pathways align with single-molecule experiments or time-resolved spectroscopy will determine real-world impact.

Still, the implications are vast. Imagine accelerating drug discovery by reliably simulating how a candidate molecule binds to its target—or predicting misfolding events linked to Alzheimer’s and Parkinson’s. Or designing enzymes with tailored catalytic pathways, guided not by trial and error but by optimal control fields derived from first principles.

This isn’t just about faster simulations. It’s about changing what we believe is possible. When a process is so rare it’s effectively forbidden, we stop looking. We assume it won’t happen. But now, with tools that can illuminate the invisible, we may finally see the full landscape of molecular possibility—one controlled step at a time.