Organic crystal structure prediction is difficult in part because molecular arrangements depend on complex interactions within and between molecules. Computationally searching for plausible structures can be demanding, so methods that extract useful structural information from molecular data may help narrow the problem. One approach combines persistent homology with an ellipsoid packing model to identify selected crystal arrangements, estimate approximate molecular shapes, and predict those shapes from molecular descriptors.
The method was examined using crystals of polycyclic aromatic molecules, a class relevant to organic semiconductors. It does not reconstruct a complete crystal structure. Instead, it estimates three ellipsoid radii as a compact description of molecular shape and tests whether those radii can be learned from information about an isolated molecule.
Why represent molecules as packed ellipsoids?
Crystal packing is often considered geometrically: atoms or molecules occupy positions that form a repeating arrangement. While spherical models can describe some inorganic structures, organic molecules are less symmetrical and often have elongated or flattened shapes. An ellipsoid can represent these differences more realistically than a sphere, while still simplifying the molecule enough for geometric analysis.
In this method, molecular centroids—the central points assigned to molecules—define the arrangement being analyzed. The study focuses on arrangements that can be described as affine transformations of four cubic lattice types: primitive, base-centered, body-centered, and face-centered. An affine transformation can change a lattice’s dimensions and angles while retaining its underlying structural relationships.
This is a targeted model, not a universal description of organic crystals. It relies on a dense-packing assumption and considers a subset of possible lattice arrangements. Its value is in extracting a simplified shape estimate when those assumptions fit the crystal.
Using persistent homology to classify molecular arrangements
Recognizing lattice-like patterns in three-dimensional centroid data can be challenging. Persistent homology offers a way to summarize how the geometry of a point cloud changes across spatial scales. Conceptually, the analysis expands regions around each centroid. As those regions meet, connected components, loops, and cavities appear or disappear. Their formation and disappearance are recorded in a persistence diagram.
The researchers derived reference diagrams for the four cubic lattice types. They then transformed a crystal unit cell into a normalized cube and generated diagrams from its molecular centroids. Comparing the resulting diagrams with the reference patterns allowed the team to classify arrangements that were sufficiently similar to one of the lattice types. The comparison used a bottleneck distance, which measures the difference between matched points in two diagrams.
This shifts the analysis from directly matching three-dimensional coordinates to comparing topological summaries. The distinction matters: persistent homology is used here to characterize the arrangement of molecular centroids, rather than the positions of individual atoms.
Estimating an approximate molecular shape
Once an arrangement is assigned to a lattice type, the packing geometry provides a basis for estimating the ellipsoid around each centroid. The unit-cell transformation and the selected lattice determine the ellipsoid’s dimensions in the original crystal coordinates. These dimensions are expressed as three radii, which can be ordered from the longest axis to the shortest.
The resulting ellipsoid is an approximation of the molecule’s shape in the crystal. It does not specify the positions of the molecule’s atoms, its conformation, or the orientations needed to define a full crystal structure. It provides partial structural information that may be useful as one component of a broader prediction workflow.
For the polycyclic aromatic dataset, the estimated ellipsoid volume correlated with a molecular volume calculated from electron density, with a reported Pearson correlation coefficient of 0.897. This supports the use of ellipsoid volume as an approximate molecular-volume measure within the analyzed data; it does not establish that the method applies equally well to every molecular class.
Predicting ellipsoid radii from molecular descriptors
The study also tested whether ellipsoid radii could be predicted from a single molecule’s chemical structure. Molecules were encoded as Extended Connectivity Fingerprints (ECFPs), binary feature vectors that represent local atomic and bonding patterns. A feedforward neural network was trained to map those fingerprints to the three estimated radii.
Prediction was evaluated using four-fold cross-validation. The reported results were strongest for the longest radius, r1, whose predictions were close to the measured estimates. Predictions for the shorter radii tended to be lower than the target values when those radii were large. The reported testing error was around 0.094 nm, while the training error was around 0.04 nm.
This difference between axes is practically important. The approach appears more useful for estimating the main molecular dimension than for recovering all aspects of a molecule’s three-dimensional shape. The fingerprints also contain chemical-structure information, so the prediction reflects more than a simple estimate based on molecular size alone.
Scope and limitations for crystal prediction
Coverage is a central limitation. In the studied datasets, the four lattice types accounted for about 30% of the polycyclic aromatic hydrocarbons, 21% of the set that also included certain heteroatoms, and 7% of the broader polycyclic aromatics set. Most structures in the broadest dataset did not meet the classification criterion used in the analysis.
The method also excludes the hexagonal crystal family and does not distinguish all possible molecular orientations. Flexible molecules may not be well represented by a single ellipsoid, while structures with substituents that affect packing, including hydrogen-bonding groups, may not satisfy the dense-packing assumption. For these cases, other methods, including force-field or density functional theory calculations, may be more appropriate.
Even when the method applies, it cannot replace conformational searches or atom-level structure determination. A complete crystal structure prediction workflow must account for molecular conformations, orientations, and interactions that are outside the ellipsoid model.
Practical implications
Combining persistent homology and machine learning offers a way to turn molecular-centroid geometry and single-molecule descriptors into a compact estimate of crystal-state shape. For materials informatics teams, the key contribution is not a complete solution to organic crystal structure prediction, but a structured way to identify certain packing motifs and estimate partial shape information.
Its practical use depends on checking whether the target materials fit the lattice and packing assumptions. Where they do, the estimated radii may provide a useful input for further screening or structural analysis. Where they do not, the method’s limited coverage and lack of atom-level detail make complementary computational approaches necessary.