Crystal structure prediction, the computational challenge of determining how a molecule will arrange itself in a solid crystal given nothing but its chemical formula, has gone from a famously intractable problem to one where the correct experimental structure now routinely appears among a method’s top-ranked candidates. A recent large-scale validation across 66 drug-like molecules found the known crystal form ranked in the top 10 predictions for every single molecule tested. Yet the field is far from finished. Molecules with many flexible bonds, crystals with more than one component, and the question of which predicted structures can actually be made in a lab all remain active frontiers where new tools, particularly machine learning, are reshaping what is possible.
Why Predicting Crystal Packing Is So Hard
A molecule in solution can tumble freely, but once it crystallizes, it locks into a repeating three-dimensional pattern defined by its unit cell, the positions of every atom, and the symmetry operations that tile copies across space. The difficulty is that most molecules have many plausible ways to pack, and the energy differences between those arrangements are tiny, often less than a kilocalorie per mole. The computational task amounts to searching a vast, high-dimensional energy surface for all its low-lying valleys, each representing a possible crystal structure, and then ranking those valleys by energy to guess which one nature will choose.
One way to visualize that surface uses a tool called a disconnectivity graph, which condenses the continuous energy landscape into a tree of discrete energy minima connected by the barriers between them. Starting from a known minimum, random perturbations are generated under the constraint that the energy stays below a steadily rising threshold, mapping out which basins connect and how deep they are.1Nature / Communications Chemistry. Global analysis of the energy landscapes of molecular crystal structures by applying the threshold algorithm The picture that emerges is humbling: even for relatively simple organic molecules, the landscape is littered with hundreds or thousands of plausible minima.
Searching the Landscape
Because brute-force enumeration of every possible packing is out of the question for anything beyond the simplest systems, the field relies on global optimization algorithms to sample candidate structures intelligently. Genetic algorithms, for instance, evolve a population of trial crystal structures over many generations, selecting the lowest-energy candidates as parents and combining or mutating their structural features to produce offspring. One recent approach, DMCrystal, encodes trial structures as pairwise atomic distance matrices and uses a genetic algorithm to reconstruct crystal packings from those distances, sidestepping some of the complications of working directly with coordinates and lattice parameters.2PubMed. Distance Matrix-Based Crystal Structure Prediction Using Evolutionary Algorithms Particle swarm optimization, random sampling followed by local relaxation, and basin-hopping methods are other common strategies, and no single algorithm dominates. In practice, most modern workflows combine a fast search stage with a more accurate re-ranking stage, precisely because the initial search generates far more candidates than can be evaluated at the highest level of theory.
Getting the Energy Right
A search algorithm is only as good as the energy function steering it. For organic molecular crystals, the energy holding molecules in place comes largely from weak intermolecular forces: hydrogen bonds, dipole interactions, and van der Waals dispersion. Standard density functional theory (DFT), the workhorse of quantum chemistry, historically struggled with dispersion because those interactions arise from correlated electron fluctuations that simple functionals miss. Augmenting DFT with empirical dispersion corrections turned out to be transformative. Tests on selected organic crystals showed that dispersion energy is the dominant component of lattice energy, especially for crystals without hydrogen bonds.3Journal of Chemical Theory and Computation. Predicting Lattice Energy of Organic Crystals by Density Functional Theory with Empirically Corrected Dispersion Energy
Dispersion-corrected DFT has since been validated extensively. A study that energy-minimized 241 experimental organic crystal structures, including their unit-cell dimensions, found an average deviation of just 0.095 angstroms from experiment, a striking level of agreement that confirmed d-DFT’s reliability for routine organic crystal work.4PubMed Central. Validation of experimental molecular crystal structures with dispersion-corrected density functional theory calculations For cases where even higher accuracy is needed, quantum Monte Carlo methods and coupled-cluster calculations (CCSD(T)) offer benchmark-quality energies, and the two approaches show excellent agreement with each other.5PubMed Central. Fast and accurate quantum Monte Carlo for molecular crystals The catch is cost: these high-accuracy methods remain too expensive for screening hundreds of candidates, so they serve as reference points for calibrating faster approaches rather than as everyday tools.
Machine Learning Potentials as an Accelerant
The tension between accuracy and speed is where machine learning has made its biggest entrance. Machine learning interatomic potentials (MLIPs) are trained on databases of quantum-mechanical calculations and learn to reproduce those energies and forces at a fraction of the computational cost. Universal MLIPs, trained on broad swaths of chemical space rather than a single molecule, now perform well enough to replace DFT in the search stage for many systems. A recent study integrated universal MLIPs with crystal structure prediction to resolve the previously unknown structures of lithium phenolate, sodium cyclohexanolate, and lithium-benzimidazol-2-one, three technologically relevant metal-organic compounds that had resisted experimental structure determination.6PubMed. Tackling complexity in crystal structure determination of metal organic compounds using machine learning interatomic potentials The acceleration is not merely incremental; calculations that would have taken weeks on a computing cluster can finish in hours, opening up classes of compounds that were previously too expensive to tackle.
Generative Models and Diffusion Approaches
Beyond speeding up traditional search-and-rank workflows, machine learning has introduced an entirely different paradigm: generating crystal structures directly. Diffusion models, the same family of generative AI that powers image generators, have been adapted to learn the statistical distribution of stable crystal structures and then sample new ones from it. The challenge specific to crystals is symmetry. A crystal looks the same if you translate, rotate, or shift it by a lattice vector, and a generative model that does not respect those symmetries will produce nonsensical outputs. DiffCSP addresses this by jointly generating lattice parameters and atom coordinates through a denoising process that is equivariant to periodic symmetry operations.7NeurIPS Proceedings. Crystal Structure Prediction: Advances and Future Directions An extension, DiffCSP++, adds explicit space group constraints to give users control over what symmetry the generated crystal should obey.8arXiv. Space Group Constrained Crystal Generation
Other frameworks encode crystal structures as point clouds rather than graphs, capturing three-dimensional spatial relationships more naturally and feeding them into a diffusion backbone trained to produce synthesizable materials.9iScience. Generative design of crystal structures by point cloud representations and diffusion model These generative approaches are still young compared to physics-based search methods, but they are rapidly improving and offer a tantalizing possibility: instead of exhaustively searching an energy surface, you could learn what stable crystals “look like” and propose candidates directly.
Blind Tests as Reality Checks
Since 1999, the Cambridge Crystallographic Data Centre has organized blind tests in which research groups predict the crystal structures of undisclosed target molecules before the experimental structures are revealed. These competitions have been the field’s most honest report card. The sixth blind test, held with five target systems spanning a rigid small molecule, a polymorphic drug candidate, a chloride salt hydrate, a cocrystal, and a bulky flexible molecule, showed substantial growth in participation and significant progress in treating flexible molecules, hierarchical ranking strategies, and the use of DFT-level methods.10PubMed Central. Report on the sixth blind test of organic crystal structure prediction methods The trend across successive blind tests is clear: the fraction of targets correctly predicted has risen steadily, and the methods that once stumbled on anything beyond rigid molecules are now routinely handling molecules with many rotatable bonds.
Complementing the blind tests, a dedicated benchmarking suite called CSPBench provides 180 curated crystal structures along with an open-source codebase, making it easier for developers to compare new algorithms on a level playing field.11arXiv. CSPBench: a benchmark and critical evaluation of Crystal Structure Prediction And industrial-scale validation is catching up: a pharmaceutical workflow tested on 66 molecules placed the experimentally known form in the top 10 for every molecule, and in the top 2 for most single-polymorph cases.12Nature Communications. A robust crystal structure prediction method to support small molecule drug development with large scale validation and blind study
Polymorphism and the Role of Temperature
Most CSP studies rank candidate structures by their static lattice energy, essentially asking which packing is lowest in energy at absolute zero with no molecular motion. Real crystals, of course, exist at finite temperatures, where vibrational entropy and thermal expansion matter. Different polymorphs, distinct crystal forms of the same molecule, can swap their stability ranking as the temperature changes, which is why a form that looks most stable computationally might not be the one that crystallizes from a warm solvent.
Vibrational contributions to free energy can shift polymorph rankings meaningfully, though thermal expansion alone has a surprisingly small effect on relative stabilities.13PubMed Central. Modelling temperature-dependent properties of polymorphic organic molecular crystals To go further, molecular dynamics simulations capture anharmonic effects and large-amplitude motions that harmonic approximations miss, but they are too expensive to apply to every candidate on a CSP landscape containing hundreds of structures. A practical solution is a multistage protocol: screen the top candidates cheaply with harmonic lattice dynamics, then advance only the most promising few to full free energy analysis via molecular dynamics. Applied to tetracyanoethylene, this workflow correctly identified the observed phase stability ordering.14PubMed Central. Assessing Polymorph Stability and Phase Transitions at Finite Temperature: Integrating Crystal Structure Prediction, Lattice Dynamics, and Molecular Dynamics
What This Means for Drug Development
Polymorphism is not an academic curiosity for the pharmaceutical industry. A drug’s crystal form affects its solubility, dissolution rate, shelf life, and sometimes even its bioavailability. If a more stable polymorph appears unexpectedly during manufacturing or storage, it can render a product unsellable. CSP offers a way to map the polymorph landscape in advance, identifying forms that might exist and estimating how much more stable they are than the marketed form. For the cholesterol drug candidate Dalcetrapib, with 10 rotatable bonds making it one of the most flexible molecules ever studied computationally, CSP calculations predicted that high pressure should stabilize otherwise inaccessible polymorphs. Experiments confirmed this: a new polymorph crystallized at modest pressures and turned out to be metastable at ambient conditions, effectively derisking the late-stage development of the compound.15PubMed Central. Combined crystal structure prediction and high-pressure crystallization in rational pharmaceutical polymorph screening
More broadly, reliable computational polymorph screening reduces the chance of costly surprises. If CSP shows that the lowest-energy structure on the landscape is the form already in development, that provides a degree of confidence no amount of experimental screening can match, because experiments can only find what crystallizes under the conditions you test. CSP, by contrast, scans configurations that may never nucleate under ordinary lab conditions but could emerge under different circumstances.16International Journal of Pharmacy and Pharmaceutical Sciences. Crystal Structure Prediction in the Context of Pharmaceutical Polymorph Screening and Putative Polymorphs of Ciprofloxacin
Discovering Functional Materials Before Making Them
Drug polymorphism is about avoiding unpleasant surprises, but CSP can also drive discovery. The concept of energy-structure-function (ESF) maps merges crystal structure prediction with property calculation to chart all the possible structures a molecule could adopt alongside the physical properties each structure would exhibit. Instead of synthesizing a molecule and hoping it crystallizes into something useful, researchers can computationally survey the landscape of possibilities and pick the most promising molecule before ever entering the lab.17PubMed Central. Functional materials discovery using energy-structure-function maps
This approach has already produced concrete results. ESF maps guided the discovery of a highly porous molecular crystal with a methane deliverable capacity rivaling the best frameworks and a density of just 0.41 grams per cubic centimeter, the lowest density molecular crystal reported at the time.18PubMed. Energy-Structure-Function Maps: Cartography for Materials Discovery The principle extends to any property calculable from a predicted structure: electronic band gaps, mechanical stiffness, optical response, and so on. In effect, ESF maps turn CSP from a structure-finding exercise into a materials design platform.
Multi-Component Crystals and Surfaces
Most CSP work has focused on crystals made of a single molecular species, but real-world systems are often more complicated. Salts, cocrystals, hydrates, and solvates contain two or more distinct components, and predicting their structures adds layers of difficulty. In salt-cocrystal pairs, even the position of a single proton, whether it sits on the acid or the base, can reshape the entire landscape of plausible packings. Calculations on pyridinium carboxylate systems showed that the correct experimental structure appeared among the low-energy candidates only when the proton was assigned to the right partner, highlighting the sensitivity of multi-component CSP to chemical details that single-molecule predictions can gloss over.19PubMed. Computational prediction of salt and cocrystal structures–does a proton position matter?
Surfaces add another dimension, literally. Molecules deposited as thin films on a substrate can adopt crystal packings that differ from their bulk form, known as surface-induced polymorphs. A new computational tool called MYTHOS tackles this by building up crystalline layers one at a time on a flat surface using molecular mechanics and molecular dynamics, identifying which polymorphs form in contact with the surface and how the structure transitions toward the bulk packing farther from the interface.20Journal of Chemical Information and Modeling. MYTHOS: A Python Interface for Surface Crystal Structure Prediction of Organic Semiconductors This is directly relevant to organic electronics, where the crystal packing of a semiconductor thin film governs charge transport and device performance.
When Theory Meets the Bench
A predicted structure is only useful if it can be confirmed experimentally, and the interplay between computation and experiment has grown tighter in recent years. When single-crystal X-ray diffraction fails, whether because crystals are too small, too disordered, or too twinned, alternative techniques can pick up the slack. Microcrystal electron diffraction (microED) works on crystals thousands of times smaller than what X-ray methods require, while NMR crystallography matches solid-state NMR spectra against CSP-generated candidates to identify which predicted packing matches the real material. A study on meloxicam, a widely used anti-inflammatory drug, combined CSP with NMR crystallography and microED to track down three elusive polymorphs that conventional methods had missed.21IUCrJ. SCXRD, CSP-NMRX and microED in the quest for three elusive polymorphs of meloxicam These hybrid workflows represent the practical frontier of the field: computation generates a shortlist, and experiment confirms or refines the answer.
The Synthesizability Gap
Perhaps the sharpest unsolved problem in crystal structure prediction is that a thermodynamically stable structure is not necessarily one you can actually make. Crystallization is a kinetic process. Nucleation barriers, solvent effects, cooling rates, and impurities all influence which polymorph appears, and a computationally predicted global minimum might never nucleate under any accessible conditions. Conversely, metastable forms that look unfavorable on the energy landscape can persist indefinitely once formed.
A new line of attack uses large language models (LLMs) trained on synthesis literature to predict whether a given crystal structure is synthesizable, what synthetic method might work, and which precursors to use. The Crystal Synthesis LLM framework treats synthesizability as a classification problem separate from thermodynamic or kinetic stability, acknowledging the significant gap between the two.22PubMed Central. Accurate prediction of synthesizability and precursors of 3D crystal structures via large language models If these models mature, they could dramatically shorten the feedback loop from prediction to laboratory realization, filtering a CSP landscape not just by energy but by practical accessibility.
High-Pressure Frontiers and Superconductivity
Some of the most dramatic successes of crystal structure prediction have come under conditions no one can hold in their hand. At pressures exceeding a million atmospheres, hydrogen-rich compounds adopt structures with no analogue at ambient conditions, and CSP calculations have been essential for identifying them. Computational exploration of binary hydride phase diagrams under high pressure has uncovered phases with unexpected stoichiometries, some of which are superconducting at remarkably high temperatures.23arXiv. The Search for Superconductivity in High Pressure Hydrides The prediction of superconducting hydrogen sulfide, later confirmed experimentally to superconduct near 200 kelvin at extreme pressure, stands as one of the field’s most celebrated achievements and a proof of concept that CSP can guide discovery in regimes where intuition fails completely. In these extreme environments, experiments are extraordinarily difficult and costly, so computation does not merely complement the lab; it leads it.