Back to search

Institutional Repository of Vilnius University

Kristalų struktūrų tyrimas naudojant mašininį mokymąsi (dirbtinius neuroninius tinklus) /

Abstract

dc:description

The determination of crystal structures and their physical and chemical properties is not only a very important process in science but also highly applicable in new material design, drug discovery, electronics, manufacturing, and other industries. Crystal synthesis and other scientific experiments with crystals require thorough procedures, deep knowledge of the field, and are time-intensive. Machine learning models, with their high training and application speeds and currently achieving good to spectacular results in many diverse areas, are being employed to address the increasing demand for crystallographic knowledge. Melting point (MP) is an important property of a crystal and is difficult to predict using computational science methods and machine learning models. While it is a fundamental property of a compound, used for determining the purity and identity of a substance, a remarkably small portion, less than 3 %, of all records in the Crystallography Open Database (COD) are labeled with MPs, which are also not always correct. Moreover, it was found that the units of MPs in COD are inconsistent – MPs are written either in Kelvin or degrees Celsius, without explicitly specified. To address the issues of rarity and incorrectness of COD MPs in particular, and the MP prediction problem in general, this work employs various COD MP data management techniques and an artificial neural network (ANN) model, trained on other COD crystal descriptors to perform MP prediction. Thorough semi-automated cleaning, validation, and augmentation of crystal MP data performed in this work provided around 8,500 COD entries with a reliable MP value and correct units, together with additional metadata describing in detail the level of credibility of each such value. An automated workflow, controlled by GNU Make and using standard Unix tools, was implemented to manage every operation from obtaining COD data to training and evaluating different ANN configurations, allowing fast and efficient experimentation with MP prediction. Although there is a wide variety of different crystal descriptors of varying complexity that can be obtained from COD, this work was narrowed to a smaller scope by using only three simple descriptors: chemical formula passed as a vector, unit cell, and atomic weight. These descriptors are available for most COD data points and can be easily calculated for missing entries. Using only these descriptors for training ANNs of various configurations, mainly multilayer perceptrons, showed a clear upper limit of MP prediction. Apart from the need for a larger MP dataset, the results emphasized the need for testing both more crystal descriptors and more sophisticated methods of representing these descriptors and testing more complex ANN architectures. On the other hand, the experimentally selected best version of ANN produced MP predictions with a mean absolute error of 50 K, which proved to be useful for determining the correct units, Celsius or Kelvin, for the majority of the remaining known COD MP data points, thus bootstrapping the number of COD MPs up to around 15,000. The results of this work – increased MP dataset together with an automated workflow allowing experimentation with different COD crystal descriptors and ANN architectures – can be used for training more sophisticated machine learning models and bootstrapping more accurate MP predictions.

Degree

thesis:*
Grantor dc:publisher
Institutional Repository of Vilnius University
Year dc:date
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Giedraitis, Tomas,

Rights

dc:rights
Statement dc:rights
  • info:eu-repo/semantics/openAccess
Language dc:language
lit

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
Vilnius University
Base URL
epublications.vu.lt/oai
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Giedraitis, Tomas,. Kristalų struktūrų tyrimas naudojant mašininį mokymąsi (dirbtinius neuroninius tinklus) /. Institutional Repository of Vilnius University, 2024. https://repository.vu.lt/VU:ELABAETD210646307&prefLang=en_US