Transfer Learning Graph Representations of Molecules for pKa, 13C-NMR, and Solubility

Published: Dec. 22, 2023

We explore transfer learning models from a pre-trained graph convoluntional neural network representation of molecules, obtained SchNet, 1 to predict 13 C-NMR, pKa, and logS sol- ubility. SchNet learns molecule by associating each atom with an “embedding vector” interacts the atom-embeddings other leveraging graph- convolutional filters on their interatomic distances. molecular energy demonstrate that atomistic embeddings can then be used as transferable for wide array properties. On one hand, atomic properties such micro-pK1 we investigate two models, linear net, inputs particular (e.g. carbon) predicts local property C-NMR). solubility, size-extensive model is built using all atoms in input. For cases, qualitatively correct predictions are made relatively little training data (< 1000 points), showcasing ease which pick up important chemical patterns. The proposed successfully capture well-understood trends pK1 solu- bility. This study advances our understanding current net representations capacity applications chemistry.

Language: Английский

---

One-shot heterogeneous transfer learning from calculated crystal structures to experimentally observed materials

Computational Materials Science, Journal Year: 2024, Volume and Issue: 235, P. 112791 - 112791

Published: Jan. 20, 2024

Language: Английский

---

Multi‐Task Multi‐Fidelity Learning of Properties for Energetic Materials

Propellants Explosives Pyrotechnics, Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 23, 2024

Abstract Data science and artificial intelligence are playing an increasingly important role in the physical sciences. Unfortunately, field of energetic materials data scarcity limits accuracy even applicability ML tools. To address limitations, we compiled multi‐modal data: both experimental computational results for several properties. We find that multi‐task neural networks can learn from outperform single‐task models trained specific As expected, improvement is more significant data‐scarce These using descriptors built simple molecular information be readily applied large‐scale screening to explore multiple properties simultaneously. This approach widely applicable fields outside materials.

Language: Английский

---

Which molecular properties determine the impact sensitivity of an explosive? A machine learning quantitative investigation of nitroaromatic explosives

Published: Nov. 15, 2022

We decomposed density functional theory charge densities of 53 nitroaromatic molecules into atom-centered electric multipoles using the distributed multipole analysis that provides a detailed picture molecular electronic structure. Three multipoles, ∑▒〖Q_0 (NO_2)〗 (the nitro groups), ∑▒〖Q_1 total dipole, i.e., polarization, ∑▒〖Q_2 (C) 〗 electron delocalization C ring atoms), and number explosophore groups (#NO_2) were selected as features for comprehensive machine learning (ML) investigation. The target property was impact sensitivity h_50 (cm) values quantified by drop-weight measurements. After preliminary screening 42 ML algorithms, four based on lowest root mean square errors: Extra Trees, Random Forests, Gradient Boosting, AdaBoost. predicted having very different sensitivities algorithms are in range 19% - 28% compared to experimental data. most important properties predicting atoms polarization with averaged weights 39% 35%, followed (16%) (10%) groups. A significant result is how contribution these depends its sensitivities: sensitive explosives (h_50 up ~ 50 cm), contribute reducing h_50, intermediate ones (~ cm ≲ 100 cm) #NO_2 increasing it other two it. For highly insensitive (h_50≳ 200 all essentially These results furnish consistent basis known also can be used developing safer new ones.

Language: Английский

---