---
title: 'MolCrystalFlow: Generative Model for Crystal Packing'
url: https://www.emergentmind.com/topics/molcrystalflow
type: topic
---

# MolCrystalFlow: Generative Model for Crystal Packing

to=arxiv_search.search  天天中彩票无法് json
{"query":"all:MolCrystalFlow OR ti:MolCrystalFlow", "max_results": 10, "sort_by":"submittedDate", "sort_order":"descending"}ดลองใช้ฟรี
to=arxiv_search.search  彩神争霸app  手机天天彩票?
to=arxiv_search.search  еиҿкаادر 彩彩票娱乐  天天彩票怎么 夜夜啪  东臣  彩票天天乐  六和彩  man  手机天天中彩票  彩神争霸充值json
{"query":"all:MolCrystalFlow", "max_results": 10, "sort_by":"submittedDate", "sort_order":"descending"}
to=arxiv_search.search  天天中彩票微信 тәы  全民彩票天天json
{"query":"ti:\"Molecular Crystal Structure Prediction via Flow Matching\" OR all:\"MolCrystalFlow\"", "max_results": 10, "sort_by":"submittedDate", "sort_order":"descending"}
to=arxiv_search.search  新天天彩票 json
{"query":"ti:\"OrgFlow\" OR ti:\"CrystalFlow\" OR ti:\"LiFlow\"", "max_results": 10, "sort_by":"submittedDate", "sort_order":"descending"}
MolCrystalFlow is a flow-based generative model for molecular crystal structure prediction that directly samples plausible periodic crystal packings for a given molecular conformer. Its central design choice is to disentangle intramolecular complexity from intermolecular packing by treating each molecule as a rigid body and jointly learning the lattice matrix, the centroid positions of molecules in the unit cell, and their orientations. Centroids and orientations are represented on their native Riemannian manifolds, and the learned dynamics are implemented with a periodic E(3)-invariant graph neural network that respects the geometric symmetries of molecular crystals. The model is benchmarked on two open molecular crystal datasets and is also integrated with a universal machine-learning interatomic potential to accelerate molecular crystal structure prediction workflows [2602.16020].

## 1. Problem setting and rigid-body formulation

Molecular crystal structure prediction (CSP) asks, given a molecule’s topology and usually a 3D conformer, how identical molecules arrange in a periodic crystal lattice. The difficulty arises from the competition between intramolecular constraints, intermolecular forces such as packing, hydrogen bonding, dispersion, sterics, and electrostatics, periodic boundary conditions, and polymorphism. The Ritonavir example is used to emphasize that small differences in packing can dramatically change stability, solubility, and manufacturability. Traditional CSP pipelines are described as generate-and-rank workflows that enumerate many candidates by random or evolutionary search and then rank them with expensive energy calculations; these searches can require enormous computational effort and still miss important polymorphs [2602.16020].

MolCrystalFlow addresses this search problem by learning a generative model over the crystal packing distribution rather than over full atomistic coordinates. Each molecule is embedded as a molecular building block using an EGNN that produces an E(3)-invariant embedding from atomistic coordinates \(x=(x_1,\dots,x_N)\in\mathbb{R}^{N\times 3}\) and atomic features \(h=(h_1,\dots,h_N)\). The final building-block embedding \(\hat h_{\rm BB}\) is combined with 18 auxiliary molecular descriptors: atom count, heavy atom count, molecular weight, chirality, H-bond donor counts, H-bond acceptor counts, rotatable bonds, aromatic rings, \(\log P\), TPSA, gyration radius, asphericity, eccentricity, planarity, and principal-axis lengths. This design keeps the generative stage focused on crystal packing while retaining explicit molecular information [2602.16020].

Under the rigid-body assumption, a molecular crystal is represented as
\[
\mathcal{C}=(F,R,L),
\]
where \(L\in\mathbb{R}^{3\times 3}\) is the lattice matrix with \(\det(L)>0\), \(F\in[0,1)^{3\times n}\) contains fractional centroid coordinates for the \(n\) molecules in the unit cell, and \(R\in\mathbb{R}^{3\times 3\times n}\) stores the corresponding rotation matrices. The atomic positions inside each molecule are never regenerated; only the rigid-body placement is. This is the model’s basic factorization of molecular CSP into lattice generation and rigid-body packing [2602.16020].

## 2. Geometric variables, symmetry constraints, and frame consistency

The three variables in \(\mathcal{C}=(F,R,L)\) live on different geometries. The lattice matrix is Euclidean, the centroid coordinates live on the three-dimensional torus \(\mathbb{T}^3\) because they are fractional coordinates modulo lattice translations, and each molecular orientation is an element of \(SO(3)\). This separation is not only representational; it determines how interpolation, message passing, and symmetry handling are defined [2602.16020].

MolCrystalFlow explicitly enforces three physical symmetries of molecular crystals. First, the target conditional distribution \(p(L,F,R\mid h_{\rm BB})\) is permutation invariant over molecules in the unit cell. Second, it is periodically translation invariant, so shifting centroids by lattice vectors and wrapping them back into the unit cell does not change the structure:
\[
p\!\left(L,\tau(F+T\cdot\mathbf{1}^\top),R\right)=p(L,F,R).
\]
Third, it respects global rotational symmetry through
\[
p(LQ^\top,F,QR)=p(L,F,R), \qquad Q^\top Q=I.
\]
These invariances are built into both the manifold parameterizations and the periodic graph neural network [2602.16020].

A further technical issue is the sign ambiguity of PCA-defined molecular axes. To make local frames consistent, the model introduces a binary axis-flip state \(\chi\). The PCA axes \(U=[u_1,u_2,u_3]\) define a sign pattern from the projections of an equivariant displacement vector, and the parity of negative signs determines \(\chi\). This partitions molecules into compatible reference-frame groups and prevents spurious sign inversions during training. The same \(\chi\)-grouping is later used in optimal transport so that trajectories do not mix incompatible axis-flip states [2602.16020].

## 3. Joint flow matching on \(\mathbb{T}^3\), \(SO(3)\), and Euclidean space

MolCrystalFlow uses flow matching to learn a time-dependent vector field that transports samples from simple base distributions to the target crystal-packing distribution. Because the variables live on different manifolds, the interpolation is constructed componentwise on their native spaces rather than in a single flat Euclidean representation. For a base sample \((L_0,F_0,R_0)\) and a data sample \((L_1,F_1,R_1)\), the interpolants are
\[
L_t=(1-t)L_0+tL_1,
\]
\[
F_t=\tau\big((1-t)F_0+tF_1'\big),
\]
\[
R_t=\exp_{R_0}\!\left(t\log_{R_0}(R_1)\right).
\]
Here \(F_1'\) is the unwrapped target centroid chosen so that the toroidal geodesic is shortest, and the orientation path is the geodesic on \(SO(3)\) defined by the exponential and logarithm maps [2602.16020].

The associated conditional velocity fields are
\[
u_t(L_t\mid L_1,L_0)=\frac{L_1-L_t}{1-t}, \qquad
u_t(F_t\mid F_1,F_0)=\frac{F_1'-F_t}{1-t}, \qquad
u_t(R_t\mid R_1,R_0)=\frac{\log_{R_t}(R_1)}{1-t}.
\]
This is the core geometric contribution of the method: the learned flow follows valid geodesics on the torus and on \(SO(3)\) instead of relying on an ad hoc Euclidean embedding of periodic positions and orientations [2602.16020].

Training uses a mixed objective with denoising targets for lattice and rotation and a velocity-field loss for centroids:
\[
\mathcal{L}(\theta)=\mathbb{E}_{t,(\mathcal{C}_0,\mathcal{C}_1)} \left[ \frac{1}{(1-t)^2}\left( \lambda_L\|\hat{L}_1-L_1\|^2 + \lambda_R\|\log_{R_t}(\hat{R}_1)-\log_{R_t}(R_1)\|^2 \right) + \lambda_F\|\hat{u}_t(F_t)-u_t(F_t)\|^2 \right].
\]
The weights are
\[
\lambda_L=0.1,\qquad \lambda_R=1,\qquad \lambda_F=2,
\]
and the time denominator is clipped at \(t=0.9\) to avoid numerical instability near \(t=1\). The base distributions are geometry-aware: fractional coordinates are uniform on \([0,1)\), lattice lengths are sampled from fitted Gaussian distributions with angles sampled uniformly on \([60^\circ,120^\circ]\), and orientations use a symmetric-rotation prior in which one orientation is sampled uniformly from \(SO(3)\) and the rest are generated deterministically from symmetry relations. Training is further stabilized by \(\chi\)-grouped optimal transport, which permutes building blocks within each \(\chi\) group to minimize geodesic distances on the torus and on \(SO(3)\) [2602.16020].

## 4. MolCrystalNet architecture and sampling dynamics

The generative backbone, MolCrystalNet, is a periodic E(3)-invariant graph neural network adapted from CSPNet/DiffCSP-style architectures. Each molecular building block is represented by its invariant embedding, auxiliary molecular descriptors, and a time embedding. Message passing operates over molecules in the unit cell using symmetry-respecting relative features,
\[
m_{ij}^{(s)}= \phi_m\!\left( h_i^{(s-1)}, h_j^{(s-1)}, L^\top L, \psi_{\rm FT}(F_j-F_i), \Psi_R(R_i^\top R_j), m_{R,L}, m_{F,L} \right),
\]
followed by aggregation and node updates. Relative fractional coordinates are encoded through Fourier features \(\psi_{\rm FT}(F_j-F_i)\), which makes periodicity explicit, while relative orientations are encoded from \(R_i^\top R_j\in SO(3)\) by converting to axis-angle form and embedding the angle variables \(\omega\) and \(\rho\) with Fourier features [2602.16020].

The lattice-coupled geometric terms
\[
m_{R,L}=\frac{L^\top L R_i^\top R_j}{\|L^\top L R_i^\top R_j+\epsilon\|}, \qquad
m_{F,L}=\frac{L^\top L(F_j-F_i)}{\|L^\top L(F_j-F_i)+\epsilon\|}
\]
tie the rigid-body geometry to the current lattice metric. After message passing, the network outputs an orientation update \(\Delta R_i\), a lattice update \(\Delta L_j\), and a centroid velocity \(u_t(F_t)\). The orientation update is multiplicative, \(R_{i,t=1}=R_i\times\Delta R_i\), and the lattice update is expressed relative to the current lattice, making it equivariant in the required sense [2602.16020].

At inference time, MolCrystalFlow integrates the learned dynamics with Euler integration over 50 steps. The update rule is
\[
x_{t+\Delta t}=x_t + (1+s t)\cdot u(x_t)\cdot \Delta t,
\]
with velocity annealing scalars \(s_{u_F}\) and \(s_{u_R}\) applied to centroid and rotational dynamics. The best values are reported around
\[
s_{u_F}\in[5,13], \qquad s_{u_R}\in[1,3].
\]
After integration, the generated \(L\), \(F\), and \(R\) are combined with the fixed internal coordinates of each rigid molecular building block to reconstruct the full all-atom crystal structure [2602.16020].

## 5. Datasets, benchmarks, and accelerated CSP pipeline

MolCrystalFlow is evaluated on two open datasets derived from molecular crystal corpora. The Thurlemann molecular crystal dataset is derived from the Cambridge Structural Database and contains 11,488 filtered structures, split into 10,000 training, 738 validation, and 750 test structures, with seven elements: H, C, N, O, F, Cl, and S. The OMC25-MCF subset is selected from the open OMC25 dataset; after filtering, it contains 46,802 structures split into 42,121 training, 3,681 validation, and 1,000 test structures, chosen as one low-energy structure per conformer under a rigid-body approximation [2602.16020].

| Dataset | Size | Split |
|---|---:|---|
| Thurlemann molecular crystal dataset | 11,488 | 10,000 / 738 / 750 |
| OMC25-MCF subset | 46,802 | 42,121 / 3,681 / 1,000 |

The main baselines are MOFFlow, described as a flow-matching model for rigid-body periodic materials, and Genarris-3, a rule-based or symmetry-conditioned molecular crystal generator with rigid press optimization. On the Thurlemann dataset, MolCrystalFlow substantially outperforms MOFFlow and Genarris-3 on direct generation and achieves markedly better lattice-volume accuracy. The reported lattice-volume RMADs are \(3.86\pm 0.07\%\) for MolCrystalFlow, \(18.8\pm 0.6\%\) for MOFFlow, \(59.0\pm 0.35\%\) for raw Genarris-3, and \(10.7\pm 0.20\%\) for Genarris-3 after rigid press. The method also generates structures in about 22 ms per structure, faster than Genarris-3 on CPU, though slower than the non-equivariant MOFFlow baseline. On the larger OMC25-MCF subset, it shows an even larger advantage over MOFFlow in both structure matching and lattice-volume accuracy, which suggests that the symmetry-aware manifold formulation scales better on a more diverse dataset [2602.16020].

For blind-test-style CSP experiments, MolCrystalFlow is integrated with the universal machine-learning interatomic potential UMA-OMC. The pipeline is: generate candidate packings with MolCrystalFlow, relax and pre-rank them using UMA-OMC, select the top candidates by approximate lattice energy, and perform final DFT ranking at PBE-D3 and PBE-MBD levels. On targets from the 3rd CCDC CSP Blind Test, this combined workflow produces low-energy candidates close to experimental polymorphs for two of the three targets. COMPACK packing-similarity analysis gives good matches for targets VIII and XI, while target X remains challenging, illustrating the difficulty of exact polymorph recovery in complex packing landscapes [2602.16020].

## 6. Scope, related methods, and limitations

MolCrystalFlow occupies a specific point in the broader landscape of flow-based generative modeling for crystalline and molecular systems. In crystalline materials generation more broadly, CrystalFlow models lattice parameters, fractional atomic coordinates, and optionally atom types using Continuous Normalizing Flows trained with Conditional Flow Matching; it is designed for crystal structure prediction and inverse materials design, with conditional generation by composition, pressure, or target properties [2412.11693]. OrgFlow specializes further toward organic crystals by conditioning directly on molecular graphs, incorporating periodic boundary conditions and a bond-aware loss, and reporting a Match Rate more than 10 times higher than existing baselines while requiring fewer sampling steps for inference [2602.20195]. By contrast, MolCrystalFlow is centered on rigid-body molecular packing, with lattice, centroid, and orientation variables as the generative state [2602.16020].

Other nearby flow models address different scientific tasks. LiFlow formulates accelerated molecular dynamics for crystalline materials as conditional generation of atomic displacements and targets lithium transport in solid-state electrolytes rather than structure prediction [2410.01464]. CGFlow extends flow matching to compositional generation and couples synthesis-pathway construction with 3D molecular pose refinement for synthesizable drug design [2504.08051]. FlexiFlow generates molecular graphs together with conformational ensembles and is explicitly described as not a direct crystal-structure generator because it does not model crystal unit cells, lattice parameters, or periodic symmetry [2511.17249]. In an earlier lineage of molecular flow methods, Molecular Flow Monte Carlo introduced a continuous normalizing flow proposal mechanism for Metropolis-Hastings sampling that suppresses extremely high-energy close contacts through an inverse-square flow [1907.07353]. These works clarify a common misconception: MolCrystalFlow is neither a generic crystal generator nor a generic 3D molecular generator, but a model tailored to periodic molecular crystal packing.

A second source of ambiguity is terminological. In a separate molecular-dynamics context, the label “MolCrystalFlow” is used for a streamlined workflow to compute the solubilities of molecular and ionic crystals by matching the crystal and solution chemical potentials, using Debye-crystal and Debye-molecule references, thermodynamic integration, endpoint free-energy perturbation where needed, and the S0 method for concentration dependence [2308.10886]. That workflow concerns solubility prediction rather than generative CSP and should be distinguished from the flow-matching model of molecular crystal packing [2602.16020].

The principal limitation of MolCrystalFlow follows directly from its representation: the atomic positions inside each molecule are never regenerated; only the rigid-body placement is. This makes the problem more tractable, but it also means the method is focused on intermolecular packing rather than intramolecular conformer search. The blind-test results reinforce that exact polymorph recovery remains difficult: target X remains challenging even within the generation-relaxation-ranking pipeline. A plausible implication is that future extensions will need to combine the current manifold-aware packing model with more explicit treatment of conformational flexibility, higher-level symmetry constraints, or stronger energy-guided refinement, as already explored in adjacent literatures on organic crystal generation and molecular conformer modeling [2602.16020].

Source: https://www.emergentmind.com/topics/molcrystalflow