In statistical mechanics and computational chemistry, calculating the absolute free energy $F$ or the partition function $Z$ of a high-dimensional system is a notoriously difficult task. The partition function $Z = \int e^{-\beta U(x)} dx$ requires integrating over an entire configuration space. Because this global integral is analytically intractable for complex systems, traditional methods rely on computing differences in free energy between states. Techniques like Thermodynamic Integration or Markov Chain Monte Carlo (MCMC) require simulating a continuous chain of intermediate, often unphysical "alchemical" states to bridge a reference state and a target state. This path-dependent integration is computationally grueling, slow to converge, and prone to critical slowing down.
A rapidly growing paradigm shifts this problem from path-dependent simulation to direct, point-wise estimation using deep generative models—specifically normalizing flows and autoregressive models. By training a generative model to approximate a target Boltzmann distribution, researchers can exploit a crucial mathematical property: these models provide exact, tractable probability densities $p_\theta(x)$ for any sample. This explicit density allows us to bypass intermediate states entirely, computing absolute free energies and partition functions directly from a single trained model.
Tractable Densities as Zero-State References
To understand why generative models can compute absolute free energies, consider how a tractable probability density function changes the integration problem. If we can sample configurations $x$ from a model's distribution $p_\theta(x)$ and evaluate $p_\theta(x)$ exactly, we can rewrite the partition function integral as an expectation over the model's distribution:
$$Z = \int e^{-\beta U(x)} dx = \int \frac{e^{-\beta U(x)}}{p_\theta(x)} p_\theta(x) dx = \mathbb{E}_{x \sim p_\theta(x)} \left[ \frac{e^{-\beta U(x)}}{p_\theta(x)} \right]$$
This simple algebraic identity turns a global integration problem into a standard importance sampling expectation. The generative model effectively acts as a perfectly known, "zero-free-energy" reference state.
As Nicoli et al. (2020) demonstrate, this architecture enables the asymptotically unbiased estimation of physical observables and free energies from a single model, completely bypassing multi-stage MCMC simulations. Similarly, Jansen et al. (2021) show that normalizing flows can directly estimate the absolute free energy of lattice field theories. This provides a massive conceptual simplification over traditional MCMC methods, which are fundamentally restricted to estimating free energy differences.
Traditional Path-Dependent Integration:
[State A] ---> [Intermediate 1] ---> [Intermediate 2] ---> [State B]
(Requires simulating multiple unphysical thermodynamic states)
Direct Generative Estimation:
[Tractable Base p_θ(x)] ==================================> [Target State B]
(Evaluates absolute free energy directly via point-wise expectations)
To scale this approach to complex molecular systems, Yu et al. (2026) introduced a coarse-to-fine autoregressive framework (CARD) that treats 3D molecular structures like a language. By generating molecular configurations with exact probabilities, the model serves as a tractable reference state, allowing the calculation of absolute free energies through standard statistical estimators like the Bennett Acceptance Ratio (BAR) without requiring any intermediate alchemical pathways. This concept of using analytical boundaries is not entirely new; early work by Lücke et al. (2012) showed that combining closed-form inner-energy limits with variational approximations of posterior entropy allows for the analytical calculation of total free energy, bypassing nested numerical integrals in probabilistic models.
Flow Layers as Non-Equilibrium Work
The connection between generative modeling and statistical mechanics runs deeper than simple importance sampling. Normalizing flows transform a simple base distribution (like a Gaussian) into a complex target distribution through a sequence of invertible, differentiable layers. This mathematical transformation can be physically interpreted as a non-equilibrium thermodynamic protocol.
In non-equilibrium physics, Jarzynski's equality relates the free energy difference $\Delta F$ between two equilibrium states to the work $W$ performed during a non-equilibrium transformation connecting them:
$$\mathbb{E}[e^{-\beta W}] = e^{-\beta \Delta F}$$
Cellini et al. (2022) recognized that Stochastic Normalizing Flows (SNFs) can be viewed directly as non-equilibrium transformations. The deterministic or stochastic steps across the flow layers act exactly like the external manipulation of a physical system. The accumulated change in the log-density across the flow maps directly to the non-equilibrium "work" done on the system. By integrating Jarzynski's equality into the generative flow, the authors compute partition function ratios and absolute free energies by averaging the work accumulated across the flow's transformations. This mechanism was later leveraged by Cellini et al. (2025) to achieve high-precision numerical simulations of effective string theories on a lattice, bypassing the limitations of traditional equilibrium sampling.
This non-equilibrium transport framework has been generalized to continuous-time transformations. Du et al. (2025) introduced FEAT, a framework that treats free energy estimation as an adaptive transport problem using stochastic interpolants. By adapting the Bennett Acceptance Ratio to work with non-equilibrium trajectories, the authors provide a statistically efficient way to combine forward and backward simulations, unifying classical estimators like Free Energy Perturbation into a single machine learning framework.
Even when explicit free energy landscapes are not the final goal, this work-energy relation simplifies sampling. Schönle et al. (2025) developed a generalized Metropolis-Hastings algorithm that constructs non-equilibrium Jarzynski-Crooks paths to bridge states defined by simplified collective variables. By using the computed non-equilibrium work directly in the acceptance criterion, they achieve unbiased sampling of metastable systems without needing to explicitly map out the underlying free energy surface.
Bypassing the Path in Practice
Eliminating the need for intermediate states yields massive efficiency gains across diverse physical domains. In material science, Wirnsberger et al. (2022) demonstrated that normalizing flows can accurately estimate the free energy differences of atomic solids. Because the model maps the target distribution directly, it circumvents the complex re-weighting techniques and intermediate states that make traditional thermodynamic integration computationally prohibitive for complex materials.
Furthermore, the model's internal state can be queried directly to extract thermodynamic properties. Schopmans et al. (2024) showed that conditional normalizing flows allow for the direct extraction of free energy surfaces from the model's internal probability density functions. This eliminates separate, computationally intensive post-processing steps typically required to calculate the potential of mean force in coarse-grained molecular studies.
This point-wise evaluation capability also means a single model can generalize across thermodynamic states. As Köhler et al. (2023) demonstrated with molecular crystal structures, training a generative model on a single reference state allows for accurate free energy calculations across a wide range of temperatures, drastically reducing the total number of expensive energy evaluations.
Mitigating Mode Collapse and Critical Slowing Down
The primary vulnerability of direct free energy estimation is mode collapse. If a generative model fails to capture a specific region of high probability (a low-energy state), the importance sampling or work estimator will miss a massive contribution to the partition function. This results in severely biased free energy estimates that can go undetected.
To address this, Nicoli et al. (2022) demonstrated that a "forward estimator" can successfully detect and mitigate systematic errors caused by mode collapse. This allows for reliable point-wise thermodynamic calculations even near challenging phase transitions where models are prone to ignoring parts of the physical state space.
Mode Collapse Bias:
True Distribution: [Mode 1] (Energy Min) and [Mode 2] (Energy Min)
Collapsed Model: [Mode 1] (Captured) and [ ] (Missed!)
Result: Partition function Z is underestimated; Free energy F is highly biased.
Rather than just detecting mode collapse post-hoc, recent frameworks embed physical constraints directly into the training process to prevent it. Yu et al. (2024) introduced Force-Guided Bridge Matching (FBM), which incorporates the target thermodynamic force field directly into the generative process. By forcing the model to favor low-energy configurations during the generation phase, the framework ensures high structural validity and eliminates the need for expensive post-hoc rejection sampling.
Finally, when physical systems approach critical points, they exhibit multi-scale fluctuations that cause standard generative models to fail to converge. To overcome this, Singha et al. (2025) developed a multilevel sampling framework informed by physics-based Renormalization Group theory. By leveraging scale invariance, this hierarchical approach provides unbiased estimates of free energy in critical regimes where standard baselines fail, unlocking the reliable investigation of large-scale critical phenomena.
Key Foundational Work
Asymptotically unbiased estimation of physical observables with neural samplers (2020) — contributed 2 concept elements to this direction.
Closed-Form Entropy Limits - A Tool to Monitor Likelihood Optimization of Probabilistic Generative Models (2012) — contributed 1 concept elements to this direction.