Applications of Artificial Intelligence and Machine Learning Algorithms to Crystallization

Chemical Reviews, Год журнала: 2022, Номер 122(15), С. 13006 - 13042

Опубликована: Июнь 27, 2022

Artificial intelligence and specifically machine learning applications are nowadays used in a variety of scientific cutting-edge technologies, where they have transformative impact. Such an assembly statistical linear algebra methods making use large data sets is becoming more integrated into chemistry crystallization research workflows. This review aims to present, for the first time, holistic overview cheminformatics as novel, powerful means accelerate discovery new crystal structures, predict key properties organic crystalline materials, simulate, understand, control dynamics complex process systems, well contribute high throughput automation chemical development involving materials. We critically advances these new, rapidly emerging areas, raising awareness issues such bridging models with first-principles mechanistic models, set size, structure, quality, selection appropriate descriptors. At same we propose future at interface applied mathematics, chemistry, crystallography. Overall, this increase adoption tools by chemists scientists across industry academia.

Язык: Английский

---

Crystal structure generation with autoregressive large language modeling

Nature Communications, Год журнала: 2024, Номер 15(1)

Опубликована: Дек. 6, 2024

Abstract The generation of plausible crystal structures is often the first step in predicting structure and properties a material from its chemical composition. However, most current methods for prediction are computationally expensive, slowing pace innovation. Seeding algorithms with quality generated candidates can overcome major bottleneck. Here, we introduce CrystaLLM, methodology versatile structures, based on autoregressive large language modeling (LLM) Crystallographic Information File (CIF) format. Trained millions CIF files, CrystaLLM focuses through text. produce wide range inorganic compounds unseen training, as demonstrated by ab initio simulations. Our approach challenges conventional representations crystals, demonstrates potential LLMs learning effective models chemistry, which will lead to accelerated discovery innovation materials science.

Язык: Английский

---

First-principles calculations: Structural stability, electronic structure, optical properties and thermodynamic properties of AlBN2, Al3BN4 and AlB3N4 nitrides

Materials Science in Semiconductor Processing, Год журнала: 2023, Номер 160, С. 107400 - 107400

Опубликована: Фев. 22, 2023

Язык: Английский

---

Crystal Composition Transformer: Self‐Learning Neural Language Model for Generative and Tinkering Design of Materials

Advanced Science, Год журнала: 2024, Номер unknown

Опубликована: Авг. 5, 2024

Abstract Self‐supervised neural language models have recently achieved unprecedented success from natural processing to learning the languages of biological sequences and organic molecules. These demonstrated superior performance in generation, structure classification, functional predictions for proteins molecules with learned representations. However, most masking‐based pre‐trained are not designed generative design, their black‐box nature makes it difficult interpret design logic. Here a Blank‐filling Language Model Materials (BLMM) Crystal Transformer is proposed, network‐based probabilistic model tinkering inorganic materials. The built on blank‐filling text generation has unique advantages “materials grammars” together high‐quality interpretability, data efficiency. It can generate chemically valid materials compositions as high 89.7% charge neutrality 84.8% balanced electronegativity, which more than four eight times higher compared pseudo‐random sampling baseline. process BLMM allows recommend operations based chemistry, useful doping. applied discover set new validated using Density Functional Theory (DFT) calculations. This work thus brings unsupervised transformer artificial intelligence A user‐friendly web app been developed be accessed freely at www.materialsatlas.org/blmtinker .

Язык: Английский

---

Designing semiconductor materials and devices in the post-Moore era by tackling computational challenges with data-driven strategies

Nature Computational Science, Год журнала: 2024, Номер 4(5), С. 322 - 333

Опубликована: Май 23, 2024

Язык: Английский

---

Towards quantitative evaluation of crystal structure prediction performance

Computational Materials Science, Год журнала: 2024, Номер 235, С. 112802 - 112802

Опубликована: Янв. 23, 2024

Язык: Английский

---

Composition Based Oxidation State Prediction of Materials Using Deep Learning Language Models

Advanced Science, Год журнала: 2023, Номер 10(28)

Опубликована: Авг. 7, 2023

Oxidation states (OS) are the charges on atoms due to electrons gained or lost upon applying an ionic approximation their bonds. As a fundamental property, OS has been widely used in charge-neutrality verification, crystal structure determination, and reaction estimation. Currently, only heuristic rules exist for guessing oxidation of given compound with many exceptions. Recent work developed machine learning models based structural features predicting metal ions. However, composition-based state prediction still remains elusive so far, which significant implications discovery new materials structures have not determined. This proposes novel deep learning-based BERT transformer language model BERTOS all elements inorganic compounds chemical composition. achieves 96.82% accuracy all-element benchmarked cleaned ICSD dataset 97.61% oxide materials. It is also demonstrated how it can be conduct large-scale screening hypothetical material compositions discovery.

Язык: Английский

---

Review on automated 2D material design

2D Materials, Год журнала: 2024, Номер 11(3), С. 032002 - 032002

Опубликована: Май 2, 2024

Abstract Deep learning (DL) methodologies have led to significant advancements in various domains, facilitating intricate data analysis and enhancing predictive accuracy generation quality through complex algorithms. In materials science, the extensive computational demands associated with high-throughput screening techniques such as density functional theory, coupled limitations laboratory production, present substantial challenges for material research. DL are poised alleviate these by reducing costs of simulating properties generating novel desired attributes. This comprehensive review document explores current state applications design, a particular emphasis on two-dimensional materials. The article encompasses an in-depth exploration data-driven approaches both forward inverse design within realm science.

Язык: Английский

---

Material transformers: deep learning language models for generative materials design

Machine Learning Science and Technology, Год журнала: 2022, Номер 4(1), С. 015001 - 015001

Опубликована: Дек. 21, 2022

Abstract Pre-trained transformer language models (LMs) on large unlabeled corpus have produced state-of-the-art results in natural processing, organic molecule design, and protein sequence generation. However, no such been applied to learn the composition patterns for generative design of material compositions. Here we train a series seven modern (GPT, GPT-2, GPT-Neo, GPT-J, BLMM, BART, RoBERTa) materials using expanded formulas ICSD, OQMD, Materials Projects databases. Six different datasets with/out non-charge-neutral or EB samples are used benchmark performances uncover biases Our experiments show that transformers based causal LMs can generate chemically valid compositions with as high 97.61% be charge neutral 91.22% electronegativity balanced, which has more than six times higher enrichment compared baseline pseudo-random sampling algorithm. also demonstrate generation novelty their potential new discovery is proved by capability recover leave-out materials. We find properties generated tailored training selected sets high-bandgap samples. each own preference terms running time complexity varies lot. our discover set validated density functional theory calculations. All trained code accessed freely at http://www.github.com/usccolumbia/MTransformer .

Язык: Английский

---

Deep reinforcement learning for inverse inorganic materials design

npj Computational Materials, Год журнала: 2024, Номер 10(1)

Опубликована: Дек. 19, 2024

Язык: Английский

---