Boltzmann Generators for Equilibrium Sampling

• Boltzmann Generators are generative models that transform simple latent distributions into approximations of Boltzmann measures while incorporating exact-density corrections for precise observable estimation.
• They leverage invertible neural networks, normalizing flows, autoregressive, and diffusion-based methods to bypass slow, local sampling and enable rapid exploration of configuration space.
• Applications span molecular simulations, free-energy calculations, and materials science, with ongoing research addressing challenges like mode collapse and computational bottlenecks.

Boltzmann Generators (BGs) are generative models for equilibrium statistical mechanics that learn a transformation from a simple reference distribution to an approximation of a system’s Boltzmann distribution and then use exact-density correction—typically importance reweighting or a Monte Carlo correction step—to recover equilibrium observables. In their original formulation, introduced by Noé et al., BGs were invertible deep generative neural networks designed to generate statistically independent equilibrium samples “in one shot” for many-body systems; subsequent work broadened the term to include conditional, coarse-grained, sequential, diffusion-based, and autoregressive variants that retain the same central objective of tractable likelihoods plus equilibrium correction (Noé et al., 2018, Rehman et al., 25 Jun 2026).

1. Statistical-mechanical formulation and core idea

In the canonical setting, the target equilibrium law is the Boltzmann distribution

$p(x)\propto e^{-\beta U(x)},$

or, in reduced notation,

$\mu_X(x)\propto e^{-u(x)}.$

The original BG construction introduces an invertible map between latent variables and molecular configurations,

$$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$

with exact density transformation through the Jacobian determinant. In the temperature-steerable-flow formalization, the same mechanism is written as

$$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$

so that sampling from an easy prior becomes sampling from a tractable proposal in configuration space (Noé et al., 2018, Dibak et al., 2021).

The defining claim of the BG paradigm is not merely that a neural network can emit plausible low-energy configurations, but that it can approximate an equilibrium measure while preserving likelihood information needed for statistical correction. This point is emphasized repeatedly in later expositions: BGs are proposal models for equilibrium sampling, free-energy estimation, and state discovery, rather than purely unconditional generators of molecular structures (Coretti et al., 2024).

A central motivation is the rare-event problem. In the original presentation, molecular dynamics and Monte Carlo were described as local procedures that move through configuration space in small steps and may require very long trajectories to cross free-energy barriers. BGs instead attempt to learn a global transport map that “repack[s]” physically relevant regions into a simple latent distribution, thereby bypassing the need to observe barrier crossing in real time (Noé et al., 2018).

2. Model classes and training objectives

The original BG implementation used deep invertible normalizing flows, specifically RealNVP-style invertible blocks, stacked into a deep network. Training combined three terms,

$J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$

where $J_{ML}$ is a maximum-likelihood loss on available configurations, $J_{KL}$ is an energy-based KL objective, and $J_{RC}$ is an optional reaction-coordinate loss for emphasizing transition regions (Noé et al., 2018).

The energy-based objective gave the framework its characteristic “training by energy” interpretation. In the original formulation,

$$J_{KL} = \mathbb{E}_{\mathbf{z}\sim\mu_Z}\left[u(F_{zx}(\mathbf{z};\boldsymbol{\theta})) - \log R_{zx}(\mathbf{z};\boldsymbol{\theta})\right],$$

so low-energy generation is explicitly balanced against an entropy or Jacobian term that discourages collapse onto a single basin. Later work generalized the same structure to temperature-conditioned models with forward-KL and reverse-KL losses,

$$\mathcal{L}=(1-\lambda)\mathcal{L}_{ML}+\lambda \mathcal{L}_{KL},$$

but now parameterized by thermodynamic state (Dibak et al., 2021).

As the field expanded, the architectural class widened. For protein-scale systems, one proposal used a split normalizing flow in reduced internal coordinates, with a backbone-first sequence of gated-attention rational-quadratic spline coupling blocks, followed by a staged training curriculum that transitions from maximum likelihood to $\mu_X(x)\propto e^{-u(x)}.$0-Wasserstein and reverse-KL refinement; this was explicitly presented as a remedy for the failure of “standard architectures and training strategies, such as maximum likelihood alone,” on HP35 and protein G (Kim et al., 2024). Sequential Boltzmann Generators later replaced prior equivariant continuous flows with an exactly invertible Transformer-based normalizing flow on all-atom Cartesian coordinates, emphasizing that exact likelihoods and cheap gradients are especially valuable when the proposal will be corrected by annealed inference-time transport (Tan et al., 25 Feb 2025).

The term “Boltzmann Generator” has also come to encompass non-flow proposals with exact likelihoods. Transferable BGs used continuous normalizing flows trained by conditional flow matching in Cartesian coordinates and evaluated zero-shot on unseen dipeptides (Klein et al., 2024). More radically, Autoregressive Boltzmann Generators replaced the global bijection altogether with the factorization

$\mu_X(x)\propto e^{-u(x)}.$1

arguing that exact likelihoods, not invertibility as such, are the indispensable component for BG-style importance correction (Rehman et al., 25 Jun 2026).

3. Reweighting, Monte Carlo correction, and free-energy estimation

The exactness claim for BGs rests on reweighting or accept/reject correction, not on perfect generation. In the original formulation, generated samples from $\mu_X(x)\propto e^{-u(x)}.$2 receive weights

$\mu_X(x)\propto e^{-u(x)}.$3

and equilibrium averages are estimated as

$\mu_X(x)\propto e^{-u(x)}.$4

The temperature-steerable-flow work gives the same mechanism in explicit canonical form,

$\mu_X(x)\propto e^{-u(x)}.$5

This is the standard BG recipe: learn an approximate equilibrium proposal, generate independent-ish samples cheaply, and reweight them to recover exact equilibrium estimates (Noé et al., 2018, Dibak et al., 2021).

That same identity yields free energies. In the conditional $\mu_X(x)\propto e^{-u(x)}.$6 formulation for phase diagrams, the importance weight

$\mu_X(x)\propto e^{-u(x)}.$7

leads directly to

$\mu_X(x)\propto e^{-u(x)}.$8

so the BG serves simultaneously as a sampler and a targeted free-energy perturbation map (Schebek et al., 2024). Closely related formulations were later used for solvation free energies, where a learned invertible map replaces hand-designed targeted free-energy perturbation transforms (Schebek et al., 20 Dec 2025).

A second correction route is flow-based Monte Carlo. Temperature-steerable flows defined a latent-noise proposal in augmented $\mu_X(x)\propto e^{-u(x)}.$9 space, followed by Metropolis correction with

$$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$0

Later work generalized the correction step from one-shot importance sampling to annealed inference. Sequential Boltzmann Generators transport initial flow samples toward the target with annealed Langevin dynamics while continuously updating importance weights, effectively implementing a continuous-time sequential Monte Carlo procedure (Tan et al., 25 Feb 2025). Diffusion-based annealed BGs placed the same idea in a broader annealed Monte Carlo framework, treating classic BG importance sampling as only one member of a larger family of asymptotically unbiased correction schemes (Grenioux et al., 28 Jan 2026).

4. Extensions across ensembles, thermodynamic states, and chemical space

A major line of work generalized BGs from one thermodynamic state to families of states. Temperature-steerable flows start from the observation that

$$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$1

and require the flow/prior pair to satisfy an analogous scaling condition, so that one model represents a temperature-parameterized family $$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$2. This turns the single-state BG into a multi-thermodynamic-state sampler and allows embedding the model into parallel tempering or other generalized-ensemble schemes (Dibak et al., 2021).

Pressure and volume fluctuations motivated a related extension to the isothermal-isobaric ensemble. The $$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$3 BG enlarges the generated state to include both particle coordinates and box degrees of freedom, with target density

$$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$4

This removes one fixed-box limitation of canonical BGs for ordered-phase transitions and yields a Gibbs free-energy estimator from generated samples (Leeuwen et al., 2023).

Conditional BGs went further by learning one transport map over a continuous thermodynamic manifold. In an $$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$5 Lennard-Jones system, one conditional flow $$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$6 was trained from samples at only one reference state and then used across

$$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$7

The paper reported more than $$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$8 effective sampling efficiency in large regions of $$\mathbf{z} = F_{xz}(\mathbf{x};\boldsymbol{\theta}), \qquad \mathbf{x} = F_{zx}(\mathbf{z};\boldsymbol{\theta}),$$9 space and estimated that the conditional-flow workflow reduced the required number of energy evaluations by about a factor of $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$0 compared with a traditional MD+MBAR-style phase-diagram mapping (Schebek et al., 2024).

Annealing in temperature rather than conditioning on it led to Temperature-Annealed Boltzmann Generators. TA-BG first trains at $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$1, then anneals through a sequence of temperatures down to $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$2 using importance reweighting and forward-KL retraining; the reported workflow used $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$3 annealing steps and a final fine-tuning stage at $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$4 (Schopmans et al., 31 Jan 2025).

Transferability across chemical space is another form of state generalization. Transferable BGs were proposed as the first framework that predicts zero-shot Boltzmann distributions for unseen molecules without retraining. On unseen dipeptides, the topology-aware “TBG + full” model achieved ESS $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$5 and correct configurations $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$6, whereas versions using only atom-type or backbone encodings were markedly weaker (Klein et al., 2024).

5. Applications in molecular simulation, free energies, and materials science

The original BG demonstrations covered toy metastable potentials, a solvated bistable particle dimer, and the protein BPTI. For the solvated dimer, the paper estimated that about $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$7 MD steps would be required to observe a spontaneous transition, whereas BG training cost about $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$8 energy evaluations and then yielded samples at essentially no cost afterward; for BPTI, the converged temperature-dependent free-energy difference between disconnected states was reported to be about $$x=f(z), \qquad p_X(x)=p_Z(z)\left|\det J_f(z)\right|^{-1},$$9 orders of magnitude faster than brute-force MD (Noé et al., 2018).

Protein and peptide sampling rapidly became the dominant application area. Differentiable rotamer sampling reformulated BG-like generation in internal torsional space and reported that decoupling energy and entropy losses reduced force-field evaluations by about three orders of magnitude relative to earlier BG implementations (Sha et al., 2023). A scalable internal-coordinate flow for HP35 and protein G reported ESS $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$0 for HP35 and $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$1 for protein G with its full staged method, whereas baseline and ablated models were essentially zero for the larger systems (Kim et al., 2024). TA-BG, SBG, and ArBG then pushed the peptide benchmarks further: TA-BG used up to three times fewer target energy evaluations than FAB and, for alanine hexapeptide, was described as the only method that accurately resolves the metastable states of the system (Schopmans et al., 31 Jan 2025); SBG reported the first equilibrium sampling in Cartesian coordinates of tri-, tetra-, and hexa-peptides that had been intractable for prior BGs (Tan et al., 25 Feb 2025); ArBG was subsequently presented as significantly improving over flow-based models across all benchmarks, especially on the $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$2-residue Chignolin system (Rehman et al., 25 Jun 2026).

Phase equilibria and materials thermodynamics formed a second major application domain. Conditional BGs were used to predict the solid-liquid coexistence line of a Lennard-Jones system in excellent agreement with state-of-the-art free-energy methods, with maximum deviation in melting temperature reported as less than about $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$3 (Schebek et al., 2024). Near the Lennard-Jones liquid-gas critical point, BGs reproduced the maximum in heat capacity, the maximum in compressibility, and their extension along the Widom line in a finite $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$4 system, while also revealing that the model’s efficiency metric closely traces the underlying phase boundaries (Santis et al., 5 Mar 2026). For large-scale crystalline materials, a local augmented coupling-flow BG with graph-neural-network conditioning was trained on small cells and transferred to systems well above one thousand atoms; on mW ice, the paper reported accurate reduced Helmholtz free energies to within about $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$5 per particle up to $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$6, and for $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$7 the local BG converged in about $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$8 GPU days, whereas earlier global BGs required $J = w_{ML} J_{ML} + w_{KL} J_{KL} + w_{RC} J_{RC},$9 GPU days for $J_{ML}$0 with ESS only around $J_{ML}$1 (Schebek et al., 29 Sep 2025).

Coarse-graining and free-energy perturbation further extended the scope. Coarse-Grained Boltzmann Generators move the proposal model to coarse-grained coordinates $J_{ML}$2 and reweight with a learned potential of mean force $J_{ML}$3, enabling asymptotically exact statistics in reduced space and supporting $J_{ML}$4-bead and $J_{ML}$5-bead coarse-grainings of alanine dipeptide (Chen et al., 11 Feb 2026). BGs were also adapted to solvation free energies, where learned invertible maps between solvent configurations improved configurational overlap for solute growth and solute separation problems in a Lennard-Jones solvent (Schebek et al., 20 Dec 2025).

6. Limitations, misconceptions, and current research directions

A persistent misconception is that BGs are primarily limited by the expressive power of a particular generator architecture. Several papers instead identify a deeper statistical issue: mode coverage under energy-only or reverse-KL training. In the RL Boltzmann Generator study for conformer generation, training against a Gibbs-score reward caused precision to increase while recall decreased, despite pruning-based diversity encouragement; the paper concluded that “the inadequacy of solely training against the energy is independent of the modeling modality” (Patel et al., 2022). TA-BG interpreted the same difficulty as temperature-dependent mode collapse, arguing that reverse-KL training becomes far less collapse-prone at sufficiently high temperature and then can be annealed downward (Schopmans et al., 31 Jan 2025).

Annealing itself introduced new failure modes. “Learning Boltzmann Generators via Constrained Mass Transport” argued that geometric annealing can suffer from mass teleportation and schedule-tuning sensitivity, and proposed KL and entropy constraints on successive intermediate distributions. On the ELIL tetrapeptide, it reported ESS $J_{ML}$6 for CMT versus $J_{ML}$7 for TA-BG, $J_{ML}$8 for FAB, and $J_{ML}$9 for reverse KL, with the abstract emphasizing more than $J_{KL}$0 higher ESS in the hardest cases (Klitzing et al., 21 Oct 2025). Off-policy log-dispersion regularization attacked the same problem from a data-efficiency perspective, using energy labels as a regularizer; the reported gains reached up to one order of magnitude in sample efficiency or target evaluations (Schopmans et al., 3 Feb 2026).

Another limitation concerns likelihood computation in continuous-time models. BoltzNCE observed that CNF-based BGs require Jacobian-trace integration along the ODE path,

$J_{KL}$1

and described this as a practical bottleneck for large systems. Its proposed replacement—an energy-based likelihood model trained with InfoNCE and score matching over stochastic interpolants—produced alanine-dipeptide free-energy differences comparable to exact-likelihood baselines while reducing inference time from $J_{KL}$2 h to $J_{KL}$3 h for the GVP Vector Field setting and from $J_{KL}$4 h to $J_{KL}$5 h for the GVP Endpoint setting (Aggarwal et al., 1 Jul 2025).

Diffusion-based annealed BGs identify a related but distinct bottleneck. Their meta-analysis on multimodal Gaussian mixtures found that even with a perfectly learned diffusion model, first-order stochastic denoising kernels fail systematically, whereas second-order kernels can substantially improve performance when covariance information is available. In the learned-model regime, all DM-aMC variants struggled, and the paper traced the main bottleneck to inaccurate log-density estimation rather than to the annealed correction alone (Grenioux et al., 28 Jan 2026).

The recent move beyond normalizing flows reflects these accumulated pressures. Autoregressive BGs explicitly reject the assumption that flow-based invertibility should define the field, arguing that flows face either topological constraints or expensive likelihoods, while autoregressive proposals retain exact likelihoods, permit sequential inference-time interventions, and scale with architectures already effective in LLMs. Their transferable model Robin, with $J_{KL}$6 million parameters, was reported to reduce zero-shot energy error $J_{KL}$7-$J_{KL}$8 on $J_{KL}$9-residue systems by over $J_{RC}$0 relative to the previous state of the art (Rehman et al., 25 Jun 2026). This suggests that the enduring invariant of the BG framework is not a specific network class, but the combination of tractable proposal densities, thermodynamic correction, and direct targeting of equilibrium sampling problems.