---
title: Conditional Boltzmann Generators
url: https://www.emergentmind.com/topics/conditional-boltzmann-generators
type: topic
---

# Conditional Boltzmann Generators

Searching arXiv for recent and foundational papers on conditional Boltzmann generators and related Boltzmann generator methods.
Conditional Boltzmann Generators are generative models that adapt the Boltzmann-generator program—direct sampling or reweighting of equilibrium configurations distributed according to a Boltzmann law—to settings in which the target distribution depends on auxiliary variables such as a molecular graph, thermodynamic state, class label, or observed variables. In the foundational formulation of Boltzmann Generators by Noé, Olsson, Köhler, and Wu, an invertible neural network learns a coordinate transformation between a complex configurational equilibrium distribution and a simple latent distribution, enabling one-shot equilibrium sampling and unbiased estimation through reweighting [1812.01729]. Later work conditions this transformation or the associated likelihood model on a 2D molecular graph for conformation-aware molecular pre-training [2312.13110], on temperature and pressure for phase-diagram mapping [2406.12378], and on structured inputs in conditional Boltzmann-machine families [1402.3346]. This suggests that “conditional Boltzmann generator” is best understood as a modeling pattern rather than a single architecture.

## 1. Conceptual scope and thermodynamic basis

The common substrate of conditional Boltzmann generators is the Boltzmann distribution. In molecular settings, conformations at equilibrium are modeled as
$$
p^*(C) \propto \exp(-E(C)),
$$
with \(C\) a 3D molecular conformation and \(E(C)\) its potential energy [2312.13110]. In the original Boltzmann-generator setting, the target equilibrium density is
$$
\mu_X(\mathbf{x}) = Z_X^{-1} e^{-u(\mathbf{x})},
$$
and a bijection is learned between \(\mathbf{x}\) and a simple latent \(\mathbf{z}\), usually Gaussian, so that generated samples can be reweighted by \(w_X(\mathbf{x}) = \mu_X(\mathbf{x}) / p_X(\mathbf{x})\) for unbiased thermodynamic averages [1812.01729].

Conditioning changes which member of a family of Boltzmann-distributed targets is being represented. In phase-diagram mapping, the target is the \(NPT\) equilibrium distribution at thermodynamic state \((T,P)\),
$$
q(\mathbf{x}, V|T, P) \propto \exp(-\beta [U(\mathbf{x}) + P V]),
$$
with \(\beta = 1/(k_B T)\) [2406.12378]. In molecular representation learning, the conditioning variable is a 2D graph \(G\), from which a latent vector \(z_G\) is inferred and used to model a conformation ensemble [2312.13110]. In conditional restricted Boltzmann machines, the conditioning variable is an input \(\mathbf{u}\), and the model parameterizes \(p(\mathbf{v}|\mathbf{u})\) directly [1202.3748].

A recurrent misconception is that conditioning eliminates the need for reweighting or exact density control. The original BG formalism relies on reweighting for unbiased estimation [1812.01729], and the phase-diagram conditional flow also estimates observables with importance weights because the learned conditional transformation is only approximate [2406.12378]. Conditioning specifies the target family; it does not, by itself, solve density estimation, overlap, or mode-coverage difficulties.

## 2. Formal objectives and conditional formulations

A major conditional formulation appears in molecular conformation modeling. In “Pre-training of Molecular GNNs via Conditional Boltzmann Generator” [2312.13110], the latent representation \(z_G\) is defined through a KL objective:
$$
z_G^* = \underset{z_G \in \mathbb{R}^D}{\arg\min}\; D_{KL}\left[\mathbb{P}_G(C)\;||\;p_\theta(C|z_G)\right].
$$
With conformation samples \(\{C_i\}_{i=1}^N\), this reduces to maximizing conditional marginal likelihood, and the estimator is written as
$$
\theta^* = \underset{\theta \in \mathbb{R}^d}{\arg\min}\left(-\frac{1}{N}\sum_{i=1}^N \log p_\theta(C_i|f_\theta(G))\right).
$$
The conditional generator itself is implemented with a diffusion model, and the pre-training loss is a variational evidence upper bound:
$$
\mathcal{L} = \frac{1}{N} \sum_{i = 1}^N \mathbb{E}_{t \sim \mathrm{Uniform}(1,T)} \left[ \lVert \epsilon - E_\theta(C_{i,t}, z_G, t) \rVert^2 \right].
$$

A second formulation is conditional transport across thermodynamic state space. In phase-diagram mapping, a single normalizing flow \(f\) is trained to transform samples from one reference thermodynamic state \((T_0,P_0)\) to a continuum of target states \((T,P)\) [2406.12378]. The conditioning variables are sampled during training from the desired thermodynamic domain, and the loss minimizes the average KL divergence between generated and target equilibrium distributions over that domain. The corresponding free-energy difference is estimated by
$$
\Delta f_{c_0 c} = -\log\left( \mathbb{E}_{(\mathbf{x}, V)\sim q_{c_0}} [ w(\mathbf{x}, V | c ) ] \right),
$$
where \(w\) is the importance weight induced by the conditional transformation.

A third formulation is post hoc Boltzmann reweighting of a pretrained generator. “Boltzmann Tuning of Generative Models” formalizes the problem as
$$
q_\beta = \arg\max_q \; \beta \mathbb{E}_q[f] - D_{KL}(q \| p),
$$
with closed-form solution
$$
q_\beta(x) = \frac{1}{Z(\beta)} p(x) e^{\beta f(x)}.
$$
The paper states that BTGM covers conditional generative modelling as a particular case: if \(f(x)=\log h(\ell|x)\), then
$$
q_\beta(x) \propto p(x) h(\ell|x)^\beta.
$$
This is not an end-to-end conditional BG in the narrow flow-based sense, but it is a Boltzmann-conditioned construction over an existing generator [2104.05252].

## 3. Architectural realizations

Conditional Boltzmann generators have been instantiated with diffusion models, normalizing flows, and Boltzmann-machine variants. The concrete mechanisms differ according to the structure of the conditioning variable and the representation of the physical state.

| Realization | Conditioning signal | Core mechanism |
|---|---|---|
| Boltzmann GNN [2312.13110] | 2D molecular graph \(G\) via \(z_G=f_\theta(G)\) | GTN encoder + Geodiff + SchNet predictor |
| Phase-diagram CBG [2406.12378] | Temperature \(T\), pressure \(P\), transformed volume \(V'\) | affine volume transformation + conditional coupling flow + permutation-equivariant transformers |
| NPT BG [2305.08483] | Pressure through \(\beta P V\) and box variable \(L\) | latent space of dimension \(ND+1\) + RealNVP coupling layers |
| BoltzNCE [2507.00846] | Variable \(y\) in conditional setting | emulator and EBM conditioned on \(y\), trained by InfoNCE and score matching |
| SCALLOP [2606.29110] | arbitrary context in conditional generation | few-step likelihood flow map with conditional divergence matching |

In Boltzmann GNN, the encoder is a Graph Transformer Network that maps the 2D molecular graph \(G=(V,E)\) to a latent vector \(z_G=[h_v,h_e]\), where \(h_v\) denotes atomic features and \(h_e\) denotes edge features. The conditional generative model is Geodiff, following Xu et al., but conditioned on \(z_G\) rather than directly on \(G\); its score network \(E_\theta(\cdot)\) is typically implemented by SchNet [2312.13110]. The intended effect is to learn a latent vector universal to multiple conformations of the same molecule.

In conditional phase-diagram flows, the workflow is explicitly state-conditioned. Box volume is transformed by
$$
V' = (1 + \alpha(T,P))V + \beta(T,P),
$$
with \(\alpha,\beta\) predicted by an MLP. Atomic coordinates are scaled to fractional coordinates, passed through a conditional coupling flow, and scaled back to the physical box. The coupling layers are conditioned on \((T,P)\) and \(V'\), use permutation-equivariant transformers, enforce periodic boundary conditions with circular spline transformations, and fix the center of mass [2406.12378]. This architecture is designed to treat solid and liquid phases on the same footing.

The isothermal-isobaric BG is closely related but conditions through ensemble control rather than explicit continuous context inputs. Its configuration space is \((L,s)\), where \(L\) denotes box dimensions and \(s\) denotes scaled particle deviations from a reference lattice. Latent space is augmented by one dimension to accommodate \(L\), and RealNVP coupling layers are used to produce invertible transformations under an \(NPT\) KL objective [2305.08483].

## 4. Molecular and materials applications

The molecular GNN application addresses a specific bottleneck in 4D-QSAR-style workflows. Molecules exist as three-dimensional structures and are in continuous motion in 3D Euclidean space, forming a potential energy surface; generating multiple conformations in advance is therefore desirable but computationally costly for drug and material discovery tasks. Boltzmann GNN uses an existing dataset of molecular conformations to generate a latent vector universal to multiple conformations from a 2D molecular graph, then transfers that representation to molecular property prediction [2312.13110]. On the five downstream datasets reported in the paper, the values for Boltzmann GNN are 0.8649 on Solubility (1.1k), 1.1586 on Malaria (10k), 0.6346 on Lipophilicity (4.2k), 0.5984 on BACE1 (3.6k), and 0.7257 on CTSD (1.1k) [2312.13110]. The same work states that the model has better prediction performance for molecular properties than existing pre-training methods using molecular graphs and three-dimensional molecular structures.

A distinct application is phase-diagram prediction. “Efficient mapping of phase diagrams with conditional Boltzmann Generators” trains a single normalizing flow to transform the equilibrium distribution sampled at only one reference thermodynamic state to a wide range of target temperatures and pressures [2406.12378]. In the Lennard–Jones demonstration, the system contains 180 particles, the target domain is \(T^* \in [0.6, 1.6]\) and \(P^* \in [1, 20]\), sampling efficiency exceeds 60% effective sample size over a large region and remains above 1% even far from the reference, relative free energies between phases match MBAR within 0.05% relative error, melting temperature errors are under \(0.01\,T^*\), and the reported cost is approximately 5x fewer energy evaluations than MBAR. The paper concludes that the solid-liquid coexistence line is recovered in excellent agreement with state-of-the-art free energy methods while significantly reducing the number of energy evaluations needed [2406.12378].

The isothermal-isobaric BG extends the scope of Boltzmann generation beyond the canonical ensemble. Because the box volume is part of the generated state, the model can sample equilibrium states at various pressures, as well as pressure-driven phase transitions [2305.08483]. The reported benchmarks are a 2D Lennard-Jones system and a Hemmer-Stell-like system. Energy, volume, pressure, and radial distribution functions are in good agreement with MD, although the BG distributions are narrower and more peaked because of anchoring to a reference lattice [2305.08483]. This is a conditional construction in the ensemble-control sense: pressure changes the target density and the loss, even when the conditioning is not amortized through a continuous embedding as in the phase-diagram model.

## 5. Relation to conditional Boltzmann machines and analytical conditional models

The literature also contains conditional Boltzmann models that are not flow-based generators in the modern molecular-simulation sense. Conditional Restricted Boltzmann Machines, denoted \(RBM_{n,m}^k\), define conditional probability distributions on the states of the output units given the states of the input units, parameterized by interaction weights and biases [1402.3346]. Montúfar, Ay, and Ghazi-Zahedi analyze their geometry and expressive power, proving results on representability of conditional Markov random fields, restricted-support conditionals, universal approximation, approximation error, and dimension [1402.3346]. These are conditional Boltzmann models, but they do not rely on the invertible-flow or reweighting constructions characteristic of Boltzmann Generators.

The 2012 CRBM work on structured output prediction makes the distinction operational. It argues that standard Contrastive Divergence-based learning may not be suitable for training CRBMs, then proposes HashCRBM for settings with relatively small sets of likely output configurations and CD-PercLoss for settings with greater output variability, such as image denoising or pixel labeling [1202.3748]. The emphasis is predictive conditional density over structured outputs, not one-shot equilibrium sampling in physical state space.

Out-of-equilibrium training on restricted Boltzmann machines adds another conditional route. In “Fast and Functional Structured Data Generators Rooted in Out-of-Equilibrium Physics,” a semi-supervised RBM includes a categorical label variable \(\ell\) in the visible layer, clamps that label during conditional generation, and is trained with two out-of-equilibrium gradients, one for label prediction and one for conditional generation [2307.06797]. The reported result is that OOE-trained RBMs become fast, robust conditional generators that produce high-quality label-conditioned samples after only 10 MCMC steps [2307.06797].

Riemann-Theta Boltzmann Machines provide yet another variant. Their main result is that the conditional density of a subset of visible units is given by a reparameterization of the original RTBM density; once the joint is modeled, all conditional densities can be inferred analytically without retraining [1905.11313]. This stands in contrast both to CRBMs, which are trained directly on \(p(y|x)\), and to flow-based conditional BGs, which typically learn transport maps or conditional score models.

A further distinction is methodological rather than architectural. BTGM explicitly contrasts itself with Conditional Boltzmann Generators by stating that CBGs typically define joint energy models over data and conditions, and require training, whereas BTGM operates post hoc on an existing generative model and covers conditional generative modelling as a special case [2104.05252]. The term “conditional Boltzmann generator” therefore spans at least two traditions: trained conditional energy or transport models, and Boltzmann-style post hoc conditioning of a pretrained generator.

## 6. Bottlenecks, misconceptions, and current research directions

A central bottleneck in Boltzmann generation is likelihood evaluation. Continuous normalizing flows require Jacobian trace integration, which is computationally intensive in high dimensions. BoltzNCE addresses this by learning the likelihood of the generated distribution with an energy-based model trained using noise contrastive estimation and score matching along stochastic interpolants, and the paper states that the method is directly extendable to conditional Boltzmann Generators [2507.00846]. On alanine dipeptide, the reported free-energy difference is \(4.08 \pm 0.13\) for BoltzNCE versus \(4.09 \pm 0.05\) for ECNF and \(4.10 \pm 0.26\) for umbrella sampling, while inference time drops from 9.37 hours for ECNF to 0.09 hours for BoltzNCE [2507.00846].

Few-step sampling and few-step likelihood estimation are a second active direction. SCALLOP replaces Hutchinson-based divergence training with a Hutchinson-free, vectorized conditional divergence matching objective, produces samples and their densities in a small number of function evaluations, and is described as naturally suited for both unconditional and conditional generation [2606.29110]. The reported empirical benefit is up to 10x inference speedup over the fastest baseline, with lower variance and lower training time than F2D2 [2606.29110].

Mode collapse and transport instability remain unresolved by conditioning alone. “Learning Boltzmann Generators via Constrained Mass Transport” argues that reverse-KL training is prone to mode collapse and that annealing-based methods can suffer from mass teleportation; its remedy is to constrain both KL divergence and entropy decay between successive intermediate distributions [2510.18460]. The paper reports more than 2.5x higher effective sample size while avoiding mode collapse and notes that the same constrained-transport logic could be extended to conditional Boltzmann Generators [2510.18460]. A plausible implication is that conditional BGs inherit the classical overlap problem twice: across metastable regions in configuration space and across changes in the conditioning variable.

Scalability to macromolecules is a further challenge. “Scalable Normalizing Flows Enable Boltzmann Generators for Macromolecules” shows that standard architectures and training strategies such as maximum likelihood alone fail on protein G and HP35, whereas a split-channel architecture with gated attention and a multi-stage training strategy using a 2-Wasserstein loss can model their conformational distributions [2401.04246]. Although the paper is not a conditional BG study, it is directly relevant to conditional settings in which a single model must cover multiple molecular contexts or thermodynamic conditions. A plausible implication is that architecture specialization and controlled training transitions are as important in conditional families as in unconditional macromolecular BGs.

The broadest methodological lesson is that conditionality is orthogonal to several of the hardest BG problems. It specifies which distribution is sought, but not how to secure tractable likelihoods, reliable overlap, stable transport, or physically plausible samples. Current work on conditional Boltzmann generators therefore proceeds along three axes at once: richer conditioning schemes, more scalable density estimation, and more robust optimization of multimodal equilibrium distributions [2312.13110; 2406.12378; 2507.00846; 2606.29110].

Source: https://www.emergentmind.com/topics/conditional-boltzmann-generators