MAGUS: machine learning and graph theory assisted universal structure searcher

National Science Review, Journal Year: 2023, Volume and Issue: 10(7)

Published: May 8, 2023

ABSTRACT Crystal structure predictions based on first-principles calculations have gained great success in materials science and solid state physics. However, the remaining challenges still limit their applications systems with a large number of atoms, especially complexity conformational space cost local optimizations for big systems. Here, we introduce crystal prediction method, MAGUS, evolutionary algorithm, which addresses above machine learning graph theory. Techniques used program are summarized detail benchmark tests provided. With intensive tests, demonstrate that on-the-fly machine-learning potentials can be to significantly reduce expensive calculations, decomposition theory efficiently decrease required configurations order find target structures. We also representative this method several research topics, including unexpected compounds interior planets exotic states at high pressure temperature (superionic, plastic, partially diffusive state, etc.); new functional (superhard, high-energy-density, superconducting, photoelectric materials), etc. These successful demonstrated MAGUS code help accelerate discovery interesting phenomena, as well significant value general.

Language: Английский

---

Geometric Deep Learning for Molecular Crystal Structure Prediction

Journal of Chemical Theory and Computation, Journal Year: 2023, Volume and Issue: 19(14), P. 4743 - 4756

Published: April 13, 2023

We develop and test new machine learning strategies for accelerating molecular crystal structure ranking property prediction using tools from geometric deep on graphs. Leveraging developments in graph-based the availability of large data sets, we train models density stability which are accurate, fast to evaluate, applicable molecules widely varying size composition. Our model, MolXtalNet-D, achieves state-of-the-art performance, with lower than 2% mean absolute error a diverse set. tool, MolXtalNet-S, correctly discriminates experimental samples synthetically generated fakes is further validated through analysis submissions Cambridge Structural Database Blind Tests 5 6. computationally cheap flexible enough be deployed within an existing pipeline both reduce search space score/filter candidates.

Language: Английский

---

XDM-corrected hybrid DFT with numerical atomic orbitals predicts molecular crystal lattice energies with unprecedented accuracy

Chemical Science, Journal Year: 2022, Volume and Issue: 14(5), P. 1252 - 1262

Published: Dec. 15, 2022

Molecular crystals are important for many applications, including energetic materials, organic semiconductors, and the development commercialization of pharmaceuticals. The exchange-hole dipole moment (XDM) dispersion model has shown good performance in calculation relative absolute lattice energies molecular crystals, although it traditionally been applied combination with plane-wave/pseudopotential approaches. This limited XDM to use semilocal functional approximations, which suffer from delocalization error poor quality conformational energies, systems a few hundreds atoms at most due unfavorable scaling. In this work, we combine numerical atomic orbitals, enable efficient XDM-corrected hybrid functionals crystals. We test new their ability predict X23 set 13 ice phases, latter being particularly stringent test. A composite approach using XDM-corrected, 25% based on B86bPBE achieves mean 0.48 kcal mol-1 per molecule 0.19 total compared recent diffusion Monte-Carlo data. These results make hybrids not only far more computationally than previous implementations, but also accurate density-functional methods crystal date.

Language: Английский

---

Inverse Design of Tetracene Polymorphs with Enhanced Singlet Fission Performance by Property-Based Genetic Algorithm Optimization

Chemistry of Materials, Journal Year: 2023, Volume and Issue: 35(3), P. 1373 - 1386

Published: Jan. 21, 2023

The efficiency of solar cells may be improved by using singlet fission (SF), in which one exciton splits into two triplet excitons. SF occurs molecular crystals. A molecule crystallize more than form, a phenomenon known as polymorphism. Crystal structure affect performance. In the common form tetracene, is experimentally to slightly endoergic. second, metastable polymorph tetracene has been found exhibit better Here, we conduct inverse design crystal packing genetic algorithm (GA) with fitness function tailored simultaneously optimize rate and lattice energy. property-based GA successfully generates structures predicted have higher rates provides insight motifs associated We find putative superior performance forms whose determined experimentally. energy within 1.5 kJ/mol most stable tetracene.

Language: Английский

---

Accurate and efficient polymorph energy ranking with XDM-corrected hybrid DFT

CrystEngComm, Journal Year: 2023, Volume and Issue: 25(6), P. 953 - 960

Published: Jan. 1, 2023

Pairing the XDM dispersion model with hybrid density functionals shows significant improvements in computed crystal energy landscapes for 4 of 26 compounds appearing first six blind tests structure prediction.

Language: Английский

---

Global analysis of the energy landscapes of molecular crystal structures by applying the threshold algorithm

Communications Chemistry, Journal Year: 2022, Volume and Issue: 5(1)

Published: July 28, 2022

Polymorphism in molecular crystals has important consequences for the control of materials properties and our understanding crystallization. Computational methods, including crystal structure prediction, have provided insight into polymorphism, but usually been limited to assessing relative energies structures. We describe implementation Monte Carlo threshold algorithm as a method provide an estimate energy barriers separating By sampling local minima accessible from multiple starting structures, simulations yield global picture landscapes valuable information on depth associated with present results applying four polymorphic organic crystals, examine influence space group symmetry constraints during simulations, discuss relationship between landscape intermolecular interactions crystals.

Language: Английский

---

Crystal Structure Prediction of Energetic Materials

Crystal Growth & Design, Journal Year: 2023, Volume and Issue: 23(8), P. 6149 - 6160

Published: July 17, 2023

The synthesis and experimental testing of energetic materials can be hazardous, but their many industrial military applications necessitate constant research development. We evaluate computational methods for predicting the crystal structures molecular organic crystals from structure as a first step in computationally evaluating materials, which could guide work. Crystal prediction (CSP) is evaluated on test set 10 with known structures, initially using rigid-molecule, anisotropic atom–atom force-field approach, followed by reoptimization predicted dispersion-corrected solid-state density functional theory (DFT). CSP force field was found to provide good results some molecules, whose are reproduced one lowest-energy predictions, more variable than typical other small molecules. Reoptimization DFT leads reliable demonstrating an approach that applied area discovery

Language: Английский

---

Predictive Modeling for Energetic Materials

Challenges and advances in computational chemistry and physics, Journal Year: 2025, Volume and Issue: unknown, P. 265 - 310

Published: Jan. 1, 2025

Language: Английский

---

Structure Prediction of Epitaxial Organic Interfaces with Ogre, Demonstrated for Tetracyanoquinodimethane (TCNQ) on Tetrathiafulvalene (TTF)

The Journal of Physical Chemistry C, Journal Year: 2023, Volume and Issue: 127(21), P. 10398 - 10410

Published: May 22, 2023

Highly ordered epitaxial interfaces between organic semiconductors are considered as a promising avenue for enhancing the performance of electronic devices including solar cells and transistors, thanks to their well-controlled, uniform properties high carrier mobilities. The structure functionality in inextricably linked structure. We present method prediction based on lattice matching followed by surface matching, implemented open-source Python package, Ogre. step produces domain-matched interfaces, where commensurability is achieved with different integer multiples substrate film unit cells. In step, Bayesian optimization (BO) used find interfacial distance registry film. BO objective function dispersion corrected deep neural network interatomic potentials. These shown be qualitative agreement density functional theory (DFT) regarding optimal position top ranking putative interface structures. Ogre investigate 7,7,8,8-tetracyanoquinodimethane (TCNQ) tetrathiafulvalene (TTF), whose has been probed ultraviolet photoemission spectroscopy (UPS), but had hitherto unknown [Organic Electronics 2017, 48, 371]. that TCNQ(001) TTF(100) most stable configuration, closely TCNQ(010) TTF(100). states, calculated using DFT, excellent UPS, presence an charge transfer state.

Language: Английский

---

Accelerated Organic Crystal Structure Prediction with Genetic Algorithms and Machine Learning

Journal of Chemical Theory and Computation, Journal Year: 2023, Volume and Issue: 19(24), P. 9388 - 9402

Published: Dec. 7, 2023

We present a high-throughput, end-to-end pipeline for organic crystal structure prediction (CSP)─the problem of identifying the stable structures that will form from given molecule based only on its molecular composition. Our tool uses neural network potentials to allow efficient screening and structural relaxation generated candidates. consists two distinct stages: random search, whereby candidates are randomly screened, optimization, where genetic algorithm (GA) optimizes this screened population. assess performance each stage our 21 molecules taken Cambridge Crystallographic Data Centre's CSP blind tests. show search alone yields matches ≈50% targets. then validate potential full pipeline, making use GA optimize root-mean-square deviation between experimentally derived structure. With approach, we able find ≈80% with 10–100 times smaller initial population sizes than when using search. Lastly, run an ANI model is trained small data set extracted in Structural Database, generating ≈60% By leveraging machine learning models predict energies at density functional theory level, has approach accuracy ab initio methods efficiency empirical force fields.

Language: Английский

---