Machine learning and AI

Over the past years I have been exploring how machine learning and AI can be leveraged for materials synthesis and microelectronics.

Applications to materials and manufacturing

In the context of materials, I am very interested in how we can develop new methodologies that can help accelerate materials synthesis and its transition to manufacturing.

I am exploring the following ideas:

• Developing self-driving tools that integrate in-situ characterization techniques with algorithms to help optimize materials towards a specific goal. I have focused primarily on atomic layer deposition, since it is an ideal model system to explore inorganic materials and it is heavily used in semiconductor processing.

• Exploring how we can use metrology tools to predict process optimization and process transfer across tools and processing techniques. The use of metrology tools during synthesis and manufacturing is ubiquitous in industry, and when properly design they can provide information not just about how your process is currently doing, but how it will behave in other conditions.

• Benchmarking large language models for materials synthesis and manufacturing. There is a lot of interest (with a dose of hype) in LLMs and their potential to assist on scientific research. Understanding how state of the art models perform in materials synthesis and processing tasks is something that is critical to determine their usefulness in real life. I am working with a team of domain experts and computer scientists to develop new benchmarks and develop a deeper understanding of their capabilities.

Development of novel algorithms more suitable for materials research

AI and machine learning research tends to look to algorithms that are widely applicable. When someone says that they have applied a blackbox algorithm to a problem, that’s actually a praise. It means that someone has been able to crack a common problem and that the resulting method or algorithm has tons of value.

Conversely, we domain scientists have to deal to very specific constraints or types of data. This means that sometimes we need to adapt existing algorithms or come up with something that is particularized to our needs. The main thing is not that the algorithm is novel, it is that we can now solve a problem that we used to have.

One of the challenges in science is that experiments are really expensive. Developing algorithms that can learn with a few examples and improve as data from other experiments is made available (ideally without the need to go through expensive retraining) is critical to maximize AI’s potential in scientific research. We have been borrowing ideas from many sources, from Bayesian processes to the insect brain to develop continual learning algorithms that target some of our unique challenges in materials synthesis and microelectronics.