Machine Learning speeds up discovery of new materials for industrial processes

A research initiative by researchers at Northwestern University and University of Toronto employs machine learning to create best building blocks for the assembly of framework materials for use in a targeted application.

The findings of the initiative is published in Nature Machine Intelligence. It demonstrates that the using artificial intelligence approaches can help to propose novel materials for diverse applications. The use of machine learning to separate carbon dioxide from industrial combustion process is an example. In fact, AI approaches are promising to accelerate the design process of materials.

Meanwhile, in a bid to improve segregation of chemicals in industrial processes, a team of researchers at Northwestern University and University of Toronto in collaboration with experts at the University of Ottawa and Harvard University set out to find the best reticular frameworks.

The frameworks can be viewed as tailored molecular sponges. The frameworks are formed via self-assembly of molecular building blocks put in different arrangements. Furthermore, the frameworks represent a new family of crystalline porous materials that have proven to be promising to address several technology challenges.

Demonstrated use of automated platform aided build-up of Design Frameworks

“Earlier, to build the frameworks, it involved building an automated discovery platform. The platform generates the design of various molecular frameworks, thus significantly reducing the time required to find optimal materials for use in this process,” said the lead author of the study.

In fact, in the demonstrated use of the platform, frameworks that are discovered are strongly competitive against some of best-performing materials used for the separation of CO2 till date.

Nonetheless, the unpredictable amount of time and massive trial-and-error efforts needed to find new materials are some perennial challenges for addressing CO2 separation.

Machine Learning finds use to develop more accurate diagnostic tool for COVID-19

In a new development against the continued fight with COVID-19, a method of producing high-quality chest X-rays scans developed. These scans can be used to detect COVID-19 more accurately than currently used methods. The finding is scheduled to be demonstrated at the IEEE Big Data 2020 Conference of December 2020.

“Meanwhile, the availability of data is one of the most important elements of machine learning. By means of this, the research has taken a theoretical step forward for generating data using MTT-GAN,” explains the lead researcher.

In fact, the need for rapid and precise testing of COVID-19 is high, which includes testing to determine if COVID-19 is impacting the respiratory system of a patient. Currently, X-ray technology is used by many clinicians to distinguish scans of possible cases of COVID-19, however, data available is limited, and thus makes it challenging to accurately classify these images.

New tool superior than currently used one

Interestingly, the tool developed by the team of researchers at the University of Maryland is an extension of generative adversarial networks. Elaborately, generative adversarial networks are machine learning frameworks that quickly generate new data based on statistics from a training set. For extending this, the team uses more advanced method what is called Mean Teacher combined with Transfer Generative Adversarial Networks.

Following this, the lead research explains why Mean Teacher combined with Transfer Generative Adversarial Networks is superior to generative adversarial networks. This is because the former generates images that are much more similar to original scans produced by X-ray machines.

Furthermore, the MTT-GAN system has the potential for improving the accuracy of COVID-19 classifying scans. This makes it an important diagnostic instrument for physicians who are still striving to understand in the number of ways this complex disease is present in patients.

Study Uses Machine Learning to Predict Performance of New Material

Defense system is one of the key factors that contribute in the progress of any country. There are many organizations such as Lawrence Livermore National Laboratory (LLNL) supporting nations to make defense systems stronger. LLNL is engaged in research and development activities. The key task of this organization is to discover and check usability of new materials required in defense activities. However, discovering a material and its actual deployment is a tedious task and may take years. Researchers at LLNL have discovered a new technique that might reduce this timing.

Deploying Advance Technology to Accelerate Deployment Process

Scientists stated that they have developed a technique that uses machine learning to aid in accelerating the development cycle. In turn, it helps in reducing time required for actual deployment of the new material. This research is accessible in the journal Materials and Design. In this research, the team focused on predicting properties of important material such as TATB—which has significant use in defense system—using machine learning.  They used combination of computer vision and machine learning, which use scanning electron microscopy (SEM) images. This helped them to avoid actual fabrication and testing part.

Scientists proved the possibility of training machine learning model to predict the material performance on the basis of SEM images. The technique offers 24% error reductions as compared to the present key techniques, which include instrument data and domain-expert assessment. Brian Gallagher is the lead author of this study. He stated, “Our motive is not only to precisely predict the performance of material, but to offer feedback to experimentalists to modify synthesis conditions to give higher-performance materials. These outcomes move us one step nearer to that motive.” Moreover, Yong Han, the corresponding author of this study, added, “Our work shows the usefulness of applying new machine-learning tactics to deal with critical materials science issues. We aim to expand on this work and deal with explainability, data sparsity, uncertainty, and development of domain-aware model.”