Energy-Guided Flow Matching Enables Few-Step Conformer Generation and Ground-State Identification
Guikun Xu, Xiaohan Yi, Peilin Zhao, Yatao Bian, [5pt] School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China · 2025
Read Energy-Guided Flow Matching Enables Few-Step Conformer Generation and Ground-State Identification by Guikun Xu, Xiaohan Yi, Peilin Zhao, Yatao Bian, [5pt] School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-22597.txt.
Introduction
Guikun Xu, Xiaohan Yi, Ziqiao Meng, Peilin Zhao, and Yatao Bian introduce EnFlow, a generative system that hunts for molecules' lowest-energy shapes. Finding ground-state conformations is central to computational chemistry, but mapping the energy landscape with physics-based pipelines is expensive. Generative models produce variety without reliable energy sense, while deterministic predictors return one structure and miss the ensemble. EnFlow couples the two approaches.
Flow-based generation is guided by an explicitly learned energy model, steering sampling toward low-energy regions and enabling energy-based ranking of what it generates. On the GEOM-QM9 and GEOM-Drugs benchmarks, the system performed strongly while needing only one or two sampling steps. Independent quantum-chemistry spot checks confirmed its learned energy scores rank conformations in a physically meaningful order.
Catalog record
| CREATOR | Guikun Xu, Xiaohan Yi, Peilin Zhao, Yatao Bian, [5pt] School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 99 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-22597
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Energy-Guided Flow Matching Enables Few-Step Conformer Generation and Ground-State Identification Guikun Xu, Xiaohan Yi, Peilin Zhao, Yatao Bian, [5pt] School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China Abstract Generating low-energy conformer ensembles and identifying ground-state conformations from molecular graphs remain computationally demanding with physics-based pipelines. Current learning-based approaches often suffer from a fragmented paradigm: generative models capture diversity but lack reliable energy calibration, whereas deterministic predictors target a single structure and fail to represent ensemble variability. Here we present EnFlow , a unified framework that couples flow matching (FM) with an explicitly learned energy model through an energy-guided sampling scheme defined along a non-Gaussian FM path. By incorporating energy-gradient guidance during sampling, our method steers trajectories toward lower-energy regions, substantially improving conformational fidelity, particularly in the few-step regime. The learned energy function further enables efficient energy-based ranking of generated ensembles for accurate ground-state identification. Extensive experiments on GEOM-QM9 and GEOM-Drugs demonstrate that EnFlow simultaneously improves generation metrics with 1–2 ODE-steps and reduces ground-state prediction errors compared with state-of-the-art methods. 1 Introduction Given a molecular topological graph specifying atom and bond types, the three-dimensional coordinates of its atoms define the set of possible molecular conformations. These conformations are essential for interpreting chemical properties and biological functions 20 ; 57 . Identifying energetically favorable conformations is a central problem in computational chemistry and drug discovery 23 . Conventional approaches, including molecular dynamics simulations 5 ; 11 ; 48 and density functional theory 47 optimizations, are computationally intensive and time-consuming, which limits their applicability to large-scale studies. Among low-energy structures, the ground-state conformation is the global minimum of the potential energy surface 46 and thus the most thermodynamically stable. It plays a key role in determining binding affinities, reaction mechanisms, and spectroscopic properties. Therefore, its accurate identification is essential for reliable prediction and rational molecular design. Figure 1 : (a) Potential energy landscape of molecular conformations: low-energy conformers populate local minima, and the ground-state conformation corresponds to the global minimum. (b) Existing generative methods can sample diverse low-energy conformers but struggle to reliably identify the ground state, whereas deterministic methods predict a single structure near the ground state and fail to capture the ensemble variability. (c) Our EnFlow framework unifies both capabilities: it efficiently generates low-energy conformational ensembles and accurately identifies the ground-state conformation by leveraging a learned, explicit energy function that simultaneously guides the generative process. Recent years have witnessed concerted efforts to harness advances in machine learning 84 ; 1 ; 28 , particularly deep generative modeling 8 ; 18 ; 60 ; 63 ; 25 ; 40 ; 42 , to address conformer generation with substantially improved computational efficiency ( 60 ; 63 ; 25 ; 40 ; 42 ; 53 ) . Although early learning-based approaches often underperformed ( 59 ; 76 ; 77 ; 58 ; 17 ; 19 ) , GeoDiff ( 78 ) introduced a principled equivariant diffusion 60 ; 25 ; 67 ; 55 formulation for molecular structures. Building on this foundation, flow matching 40 has emerged as an alternative to diffusion models 25 and yields superior sample efficiency and faster generation ( 40 ; 42 ) . Subsequent studies further refine flow-matching pipelines to attain a more favorable balance between conformational fidelity and computational cost ( 22 ; 7 ) . Despite substantial progress in identifying low-energy conformations for a given molecular graph, pinpointing the global minimum, the ground-state conformation, remains challenging. To address this challenge, recent studies employ deterministic prediction methods that model the ground state directly, which reduces reliance on stochastic sampling and subsequent ranking. In these approaches ( 79 ; 74 ; 44 ; 71 ; 32 ) , deterministic neural networks, particularly Graph Transformers 80 with task-specific modifications, are trained to regress the ground-state conformation from datasets that provide a single ground-state label for each molecule. As shown in Fig. 1 (b), a practical method that jointly models the low-energy conformational ensemble and the specific ground state within a single unified framework remains lacking. Generative approaches 78 ; 22 can sample diverse structures but do not provide an explicit, calibrated view of the energy landscape, making ground-state identification unreliable. Prediction approaches 74 ; 32 , by contrast, predict a single structure intended to approximate the ground state and therefore fail to capture ensemble variability or quantify uncertainty. To address the above challenges, we propose EnFlow , as illustrated in Fig. 1 (c), a unified framework for generating low-energy molecular conformational ensembles, together with a demonstrably fast learning-based certification of the ground-state conformation. Fig. 2 provides an overview of EnFlow: (1) An energy-guided FM framework is proposed to enable principled guidance along non-Gaussian FM paths, providing a unified solution to the conformation problems described above. Unlike conventional guidance techniques developed for Gaussian-path FM and Diffusion models ( 13 ; 26 ; 43 ) , such approaches do not directly extend to non-Gaussian settings. In conformation modeling, the prior distribution often deviates substantially from a Gaussian prior (e.g., the Harmonic Prior 65 ), reflecting strong domain-specific structural inductive biases that are essential for capturing molecular structure and consequently inducing flow dynamics that differ fundamentally from the Gaussian-path regime. (2) An energy-based guidance mechanism is formulated to incorporate energy information directly into the flow dynamics. Samples are thereby steered toward lower-energy regions, leading to substantially improved sample fidelity, particularly under few-step ODE sampling. (3) A central component of the framework is an explicit energy function that models the conformational energy landscape. This function is parameterized using an energy-based model (EBM) 38 , trained via the recently proposed Energy Matching objective ( 4 ) . Beyond capturing the global energy structure, the EBM is further fine-tuned to match empirical per-molecule energy distributions, enabling the resolution of subtle energy variations across different conformations of the same molecule. (4) Beyond improved ensemble generation, coupling a strong generator with a well-aligned EBM endows EnFlow with an explicit and calibrated energy function. This enables efficient ground-state identification either by ranking generated conformations according to predicted energies or by direct one-shot generation that produces a single conformation per molecule using the proposed energy-guided sampling procedure. In this work, the proposed framework is instantiated using ET-Flow ( 22 ) with a Harmonic Prior 65 as the generative backbone, transporting harmonic samples to physically plausible molecular conformations. Supported by empirical experiments on GEOM-QM9 and GEOM-Drugs, EnFlow achieves SOTA performance in both small-scale and large-scale settings for 1–2 ODE-steps conformational ensemble generation and for ground-state conformation prediction, significantly surpassing existing methods. By acting as a fast, differentiable surrogate for quantum mechanics, EnFlow paves the way for biologically aware and thermodynamically consistent molecular design. Figure 2 : EnFlow illustrations. (a) The energy-guided flow matching framework, in which an EBM trained via the Energy Matching technique provides guidance during the flow matching process. (b) Architectural overview of the vector field and the energy model. For comparision fairness, we used the same backbone architecture as that of ET-Flow 22 (c) Illustration of improved one-step ODE sampling achieved through energy-guided sampling. (d) Using the well-trained energy function to certify the ground-state conformation. 2 Results While prior methods typically address conformation ensemble generation or ground-state certification in isolation, our approach unifies these two tasks within