Manifold Diffusion for Structure Generation of Transition Metal Complexes
Luca Schaufelberger, Kjell Jorner · 2026
Read Manifold Diffusion for Structure Generation of Transition Metal Complexes by Luca Schaufelberger, Kjell Jorner (2026) 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-2606-00666.txt.
Introduction
Schaufelberger and Jorner build a tool for chemistry in this 2026 paper. Transition metal complexes sit at the heart of catalysts, drugs, and new materials. What they do depends a lot on their three-dimensional shape. Yet their odd bonds and wide range of electron states make shapes hard to generate well.
Their model, TMCgen, grows shapes with a diffusion process set on a curved space. It works on the angles between the metal and its ligands, plus twists and turns of the ligands. That focus keeps the search on the traits that matter most. The authors report that TMCgen makes sound coordination shapes for a broad set of real complexes. It needs few steps, so it runs fast. The work points toward design of new complexes tuned for a chosen trait.
Catalog record
| CREATOR | Luca Schaufelberger, Kjell Jorner |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.00666
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Manifold Diffusion for Structure Generation of Transition Metal Complexes Luca Schaufelberger, Kjell Jorner Abstract Transition metal complexes are central to catalysis, drug design, and materials science, with relevant properties strongly sensitive to their three-dimensional geometry. However, the electronic diversity and unconventional bonding environments of transition metal complexes pose a major challenge for accurate structure generation. In this work, we introduce TMCgen, a manifold diffusion machine learning model that efficiently and accurately generates geometries of transition metal complexes. By formulating the diffusion process over the metal-ligand coordination angles, combined with torsional and rotational diffusion of the ligands, TMCgen focuses on the key geometric degrees of freedom of transition metal complexes. TMCgen shows strong performance in generating accurate coordination environments on a diverse set of experimentally derived bioinorganic and organometallic complexes while requiring only few inference steps, enabling efficient generation. Our results demonstrate the potential of manifold-based generative modeling for data-efficient geometry generation, paving the way for property-conditioned design of transition metal complexes. keywords Generative machine learning, transition metal complexes, manifold diffusion, coordination environments 1 Introduction Figure 1: Overview of this work. (a) Left: Examples of transition metal complexes used as catalysts, drugs and functional materials, with the central metal atom and the ligating atoms highlighted. Right: Simplified depiction of two exemplary coordination environments, tetrahedral and square planar (b) Challenges addressed by (c) the manifold diffusion model TMCgen introduced in this work. Transition metal complexes (TMCs) play a crucial role in catalysis 19 , drug design 35 and materials science 27 due to their unique structural diversity and electronic properties. Monometallic TMCs consist of a central transition metal atom bonded to surrounding molecules or ions, called ligands (see Fig. 1 a). In catalysis, transition metal complexes are crucial to achieving important transformations such as C–H activation 13 , asymmetric hydrogenation 50 , and cross-coupling reactions 23 , many of which underpin sustainable synthesis and green energy storage technologies 6 . Transition metal complexes have also found broad application as drugs, where they have demonstrated inhibitory effects on various proteins 4 ; 41 ; 15 . The tunable electronic structure of transition metal complexes enables their use in functional materials, for example in spintronics 10 and molecular electronics 1 . The unique properties of TMCs are determined by the three-dimensional arrangement of ligands surrounding the metal center. Specifically, both the metal coordination environment, i.e. , the angular distribution of ligands around the metal center ( e.g. , tetrahedral or square planar, see Fig. 1 a), and the relative positioning of ligands ( e.g. , cis or trans ) strongly affect their electronic structure. For example, for \text{Ni}{\vphantom{\text{X}}}^{\text{2+}} ( d^{8} ) complexes, strong-field ligands favor square planar geometries ( e.g. , \text{[}\text{Ni}\text{(}\text{CN}\text{)}\text{}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{]}\text{}{\vphantom{\text{X}}}^{\text{2\hskip 0.90417pt--\hskip 0.90417pt}} ), which are low-spin and diamagnetic, whereas weak-field ligands favor tetrahedral geometries ( e.g. , \text{[}\text{NiCl}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{]}\text{}{\vphantom{\text{X}}}^{\text{2\hskip 0.90417pt--\hskip 0.90417pt}} ), which are high-spin with two unpaired electrons and therefore paramagnetic 19 . To achieve a large-scale computational design of transition metal complexes 45 , accurate generation of their three-dimensional geometries is essential. However, these geometries are heavily modulated by steric and electronic ligand effects that cause them to often deviate significantly from idealized coordination environments 28 . For instance, bulky ligands tend to adopt positions that minimize steric hindrance, while multidentate ligands — those with multiple metal-binding sites — can impose rigid geometries due to the constrained bite angle. Due to these complexities, generating realistic coordination geometries for TMCs remains highly challenging. Traditional approaches to 3D-geometry generation include metadynamics methods 36 , which however have limited usability in large-scale discovery campaigns due to their computational expense. Knowledge-based cheminformatics methods such as RDKit’s ETKDG algorithm 38 have been primarily designed for generating conformers of organic molecules and exhibit limited applicability for TMCs, as they do not incorporate experimentally derived coordination-angle preferences for TMCs analogously to how torsional preferences are incorporated for organic molecules. Instead, they use default coordination environments and position ligands randomly on the available coordination sites if this has not been defined a priori in the input. Among approaches that produce transition metal complex geometries 43 ; 46 ; 7 ; 34 ; 47 ; 8 , molSimplify 22 , which combines cheminformatics methods with classical force fields, has been widely used for automatic screening of TMCs. However, it requires prior input of the coordination environment, and either places ligands randomly at coordination sites or based on explicit input, limiting its effectiveness for targeted design — particularly since TMCs often possess numerous stereoisomers. Generative models such as those based on diffusion have recently surpassed the performance of cheminformatics tools for 3D-geometry generation in terms of precision and coverage of the conformational space 26 ; 48 , while being sufficiently fast to be used in large-scale discovery campaigns. They further allow for conditional geometry generation based on desired target properties 20 ; 49 , enabling inverse design workflows wherein molecular geometries are generated to meet specific functional or electronic criteria. Generative models have also been used to design new ligands or whole transition metal complexes 25 ; 9 ; 45 . In chemistry and biology, diffusion models fall into two categories: Euclidean diffusion models, which operate directly on 3D atomic positions, and manifold diffusion models, which operate on a lower-dimensional manifold, i.e. , a chemically-relevant coordinate space such as torsion angles 26 ; 11 ; 53 ; 51 ; 40 . Euclidean diffusion models provide flexibility but rely on large and diverse training datasets, which are not available in many target domains such as transition metal complexes. For instance, while the GEOM dataset for small organic molecules 2 contains 37 million conformers, the experimentally derived TMC dataset 30 used in this work comprises only 61 thousand structures, i.e. , nearly three orders of magnitude fewer. On the other hand, manifold generative models incorporate domain-specific information by limiting diffusion to the most chemically meaningful degrees of freedom, leading to data-efficient learning and providing guarantees for chemical validity, e.g. reasonable bond lengths 26 ; 11 ; 53 ; 16 ; 40 . For example, diffusion over torsion angles exhibits strong performance by reducing the degrees of freedom, which are the most important for most small organic molecules. Recently, the PuckerFlow model introduced by us has demonstrated the strength of manifold generative models in the accurate generation of cyclic conformers 40 . Manifold generative models have also been shown to allow for sampling with few inference steps, enabling compute-efficient generation 26 . Despite these advances, current manifold diffusion models are limited to organic molecules, as they address insufficient degrees of freedom to generate the coordination environments of transition metal complexes. For example, the approach used in torsional diffusion keeps all bond lengths and angles fixed, including the coordination bonds and angles of ligands around a metal center. Therefore, new methodologies are required that can model the coordination environments of transition metal complexes effectively. Here, we introduce TMCgen, a manifold diffusion model that achieves strong performance in generating three-dimensional geometries of transition metal complexes. We demonstrate the performance of TMCgen on a highly diverse set of transition metal complexes extracted from experimental crystal structures, on which our model generates geometries with accurate quantum-mechanical properties. By constraining the diffusion process over the relevant manifold, i.e. ,