ML assisted design of experiments for solid state electrolyte

Machine Learning (ML) has become a powerful tool in experimental design, especially in the optimization of complex material synthesis processes. One popular approach is Bayesian optimization, which is used to efficiently explore and optimize experimental parameters. Bayesian optimization is particularly useful in scenarios where experiments are costly or time-consuming, as it strategically selects the most promising conditions to test next based on prior data. By continuously updating its model with new experimental results, Bayesian optimization helps identify optimal conditions with fewer trials, accelerating the discovery of materials with desired properties.

Batteries with solid electrolytes are safer, offer higher energy density, and last longer, making them a promising alternative to traditional liquid electrolyte batteries. To optimize its performance, we need to systematically study how different experimental factors affect its design. In this study, we used ML to guide the design of experimental conditions for synthesizing one solid-state electrolyte material and used Kadi4Mat to efficiently manage the research data during the experiments. We focused on factors like variations in precursor concentrations, sintering temperature, and holding time. We compared different ML models built from previous lab data to find the best one for designing new experiments. This approach led to the discovery of a good sample with competing performance after just a few iterations. To understand why these samples have high ionic conductivity, we analyzed their phase compositions and crystal structures using X-ray diffraction and examined their microstructures with scanning electron microscopy. Our results show that applying machine learning to design experimental conditions can help researchers create desired materials more efficiently, reducing the number of experiments needed.

More information can be found in our publication: DOI: 10.3389/fmats.2022.821817

LLM-based chatbot

The rapid advancement of battery technologies has generated a vast amount of research data and scientific literature. Manually analyzing this wealth of information is a daunting and time-consuming task, especially when trying to keep up with the latest findings. Natural Language Processing (NLP) techniques, particularly those powered by Large Language Models (LLMs), offer a powerful solution to this challenge. LLMs can process and understand large amounts of text, allowing researchers to efficiently extract key insights from complex battery research data and publications. These models can summarize long articles, identify relevant studies, and even extract specific data points or trends, helping researchers quickly gather and synthesize the information they need to drive innovation.

By applying NLP techniques to battery research, scientists can automate the extraction of valuable information from diverse sources, including academic papers, patents, and technical reports. NLP can help identify correlations between different research findings, identify gaps in the literature, and uncover emerging trends. This not only speeds up the research process, but also provides a more comprehensive understanding of the field. As battery technologies continue to evolve, the integration of NLP tools into the research workflow will become increasingly important, enabling researchers to navigate the ever-growing body of knowledge and make more informed decisions in the development of next-generation energy storage solutions.

More information can be found in our LLM-based chatbot demo: https://huggingface.co/spaces/Kadi-IAM/LISA

Generating digital twins of microstructures using ML to accelerate battery simulation studies

Deep Learning (DL) is revolutionizing the field of materials science by accelerating simulations and enabling the generation of realistic microstructures. By using DL models, researchers can significantly speed up the process of simulating complex material behavior, reducing the time and computational resources required. In addition, generative Artificial Intelligence (AI) models, such as those used to create realistic images, can now be applied to generate microstructures that closely resemble real materials. These advances make it possible to more efficiently explore material properties, optimize designs, and predict performance without relying solely on traditional experimental methods.

Porous structures are commonly found in battery materials. A thorough study of the properties of these structures is important to analyze their corresponding influence on the battery. A promising approach is to generate digital twins of real structures, validate them with real experiments, and then use data-driven methods such as artificial neural networks to understand the relationship between the structure, the manufacturing process, and the material properties. In this study, we use a Variational AutoEncoder (VAE) neural network to capture and simplify the 3D structure of porous materials. This allows us to model the relationship between the structure and its properties, and also to predict how changes in the manufacturing process might affect the material. Our method provides a reliable way to learn about these structures without the need for labeled data, enabling a deeper understanding of porous materials.

More information can be found in our publication: DOI: 10.1016/j.actamat.2023.118922

Enhancing analysis of spectral data with ML models

In materials science, the characterization of complex materials often generates large and complex data sets that can be difficult to analyze using traditional methods. Machine Learning (ML) has emerged as a powerful tool for efficiently handling and interpreting such data. Using advanced algorithms, ML can identify patterns, make predictions, and extract meaningful insights from large and complex datasets that would otherwise require extensive manual analysis. This approach not only speeds up the research process, but also increases the accuracy and depth of analysis, allowing researchers to uncover subtle relationships and characteristics within the data. In this context, ML offers significant potential to transform materials research by simplifying the interpretation of complicated characterization data and opening up new avenues for discovery and innovation.

A technique called Time-of-Flight Secondary Ion Mass Spectrometry (ToF-SIMS) can be used to analyze surfaces of materials, but the data it generates is often complex and time-consuming to interpret manually. In this study, ML method, particularly logistic regression, is employed to classify different material components and identify key ions from different compounds. These models are then used to determine the compositions of both mixtures and real samples from their ToF-SIMS data. The ML approach successfully identifies the characteristic ions and accurately predicts the compositions of new samples. The study also discusses the practical benefits and limitations of this method, highlighting its potential to simplify the analysis of other materials.

More information can be found in our publication: DOI: 10.1021/acsami.3c09643

The use case involves the evaluation of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) data from 75 subjects from the COSYCONET study, which contains around 600 patients with chronic obstructive pulmonary disease (COPD) of different disease severity. For the successful analysis of this data breathing motion as shown in the figure below has to be identified and removed. Since manual breathing detection is cumbersome, an AI model for automatic breathing detection was developed by training and testing three different AI hypermodels.

The automatic synchronization of the AI project with Kadi4Mat results in records with metadata and files for each component (see Figure 3 and Figure 4). Due to the usage of the KadiAI ontology, the links between the single records lead to the automatic generation of a knowledge graph for the whole AI project (see Figure 5). This use-case shows the successful implementation of a concept that automates RDM for AI applications, minimizing the overhead for the individual researcher and recording rich metadata and links between entities toward FAIR data.

Figure 3

Figure 4

Figure 5