In this specific work, we show how to integrate an AI agent based on LLMs with an experimental materials synthesis tool, in this case one of our trusty ALD reactors. This is a work that was long in the making, and that would not have been possible without our indefatigable postdocs working to keep the reactor alive through the testing and upgrades.

The paper also includes some preliminary results evaluating agents based on state of the art (as of summer/fall of 2025) LLMs in materials synthesis tasks. In these challenges, the outcome of a query like “grow 20 cycles of Pd” should yield a specific instruction to run a process in the tool. While the paper focused on ALD, the methodology of how test AI agent’s ability to do useful things with experimental tools is general.

I am happy to say that no reactor was harmed during the making of this paper. However, our results show that there is still a gap in capabilities of the AI models, particularly for processes that are not well represented in the scientific literature. I hope that these results spur the development of better models and encourage others to evaluate how well AI performs in their own research settings. After all, we are scientists so quantifying how good/bad methods work is part of what we should do.

I am in the process of getting some of the relevant components online in an open source repository, hopefully this June.

In order to make sense of what’s going on, I try to place AI methods in a space defined by three different axes:

1. Technology readiness level: basically, is this something that is AI research (low TRLs) or is it ready to be used in real life (TRL 9)?

2. How specific is the model to my research field? Is this something like email, something useful but general-purpose or is this something that has developed for a specific problem?

3. What is the visibility of this AI method? And by that I mean, is it something that is likely to be cited (like a specific model), disclosed (like a tool) or acknowledged (like a facility)?

Based on this you can put AI works in this 3D space. At low TRL levels sit the vast majority of papers and projects involving AI for materials science. These works, which focus on developing or demonstrating new methods, are essentially AI research. The point is not to be immediately useful to the researcher, they are essentially methods papers on a materials science topic where the publication itself is the main outcome.

In this region, whether AI lives up to its hype is largely immaterial to the work itself, affecting primarily the shelf life of many of these publications. One may dispute whether investing a lot of resources in it is warranted, but the situation is no different than, say, any other idea going through the research hype curve. Right now, though, communities are growing in this space, as we can see in conferences, journals, and funding priorities. The range of specificity and visibility in the research of these papers is really broad.

At the other side of the spectrum, at high TRLs, sits what I call AI for research. These are the methods that are or will be actively used by researchers who are not necessarily the ones who developed them. Reusability is the key here. Some of these methods will be general (like using generative AI to accelerate the development of scientific codes), but most will be very specialized. The range of visibility is a bit skewed towards higher visibility with respect to more exploratory work. AI methods that broadly support the work of the researcher are not likely to be considered AI for research, much like email is not considered a research tool. They tend to be skewed towards higher specificity for the same reason.

The corner with high TRL, domain-specific AI is the most intriguing to me, because I think we do not know yet how impactful AI will turn out to be, how the research enterprise will shape and be shaped by AI. In fact, we as domain scientists have barely started to quantify its value and potential gains. This region is where the aspirational goals of AI meet the reality of research.

What ifs?

Once you start placing hypothetical tools in this space you can start asking questions about the potential impact of AI in our research.

For instance, as a researcher, how should I communicate useful, general-purpose, but low-visibility ways of using AI? If I use an LLM to summarize papers and classify my sputtering papers, is that something that should be published in a physics or materials science journal simply because it involves materials? Or should that become part of how things are done in a group or a lab?

Here is another example: is the application of general-purpose models to a specific scientific problem a research goal in itself or is it just a means to achieve something else, much like an electron microscope? I assume that at some point in history, using an SEM to view something was noteworthy in itself until progressively it became commonplace and the novelty wore off. It seems that right now we are still in the “I used AI for X”, but at some point we will get over that phase.

Another way of framing this question is: what areas of this space should materials science methods papers cover? Is the application of a general model to query a database of materials properties a materials science or an AI paper? Should the scope be limited to specific, high-visibility parts of the space?

Finally, how do we evaluate the performance or impact of some of the high-visibility, high TRL models being developed right now? If we take a specific AI model to extract information from X-ray data, how can we quantify how good it is? Does this type of comparison belong in a high impact journal?

In this preprint on arxiv I explored how agents based on reasoning models perform in ALD process optimization tasks. The results are both impressive and somewhat mixed. One one hand, reasoning models were able to understand and execute elementary atomic layer deposition process optimization tasks. On the other, they sometime either fail or struggle to find the optimal conditions, which makes them borderline usable for real world application.

For this work I didn’t use a finetuned model, but a commercially available model. The motivation is that these models are widely used, and therefore there is a lot of value in understanding how they perform. The other more prosaic reason is that there are limitations to the type of models that we can access. This left out some good open weight models such as DeepSeek-R1.

For the test, I designed a few ideal and non-ideal self-limited processes that representative of the type of ALD processes a researcher is likely to encounter in the wild. For each of these processes, the task was very simple: find the optimal dose time for the precursor and co-reactant that leads to a saturated growth per cycle with a process time that is as low as possible. This is a very much an idealized version of the type of optimization that is carried out over and over again.

The work is currently under review. As soon as the paper is published I will release the code required to run these challenges. Hopefully someone can take on the challenge of designing performant agents based on open weights models that can be run locally.

A detour on RIS data files

Not sure what reference manager people use these days but, if you use something like Zotero or have downloaded any reference file from Springer Nature, you most likely have come across a RIS file.

The RIS format is old, so old that I had to use the Wayback machine to find an original specification. Now I keep a copy in the gris documentation for reference.

It is also a format that publishers tend to play with, many years back WebOfScience used a RIS format with different tag formats when you downloaded references as other. Recently, AIP has messed up with the format adding headers that were not part of the original specification.

gris and its updates

The new version of gris now doesn’t break when reading files with headers that should not be there. It now has proper documentation and a bit of testing.

The main purpose of gris is to parse these old files into Python objects that can be then manipulated or converted into more universal formats, such as JSON. While it is completely unrelated to my daily work, as it happens it has made its way to some of my papers as well.

To do list

There are a few items in my to do list:

• I need to expand the collection of input files used in tests.
• I need to document the json output and demonstrate that the RIS -> JSON -> RIS roundtrip works.
• I need to document the code.
• I need to decide whether I want to add to gris a proper semantic layer.

This is an issue when trying to write code dealing with physical systems. A quick perusal of Python packages shows a zoo of possible solutions designed to help you do operations with physical magnitudes. To add to the confusion, here is a new stable release of my side project: properunits.

Intro to properunits

The core idea of properunits is that dealing with physical magnitudes should simply be an IO problem. A user (me) wants to work with a specific unit and my code wants to work with its own unit. Once you do that transformation, you should be good to go because you have transformed your 5 gallons into a number using the proper units your code can work with. In my case, this means working with SI+ units, which to me it means SI units plus maybe one or two exceptions.

So this is how it works:

from properunits import Temperature

my_temp = Temperature(20, 'C')
the_proper_temperature = my_temp.x

What properunits does is provide a consistent interface to transform an input with user-defined units into a proper value (in the example K) that my code can deal with.

You can know which unit is considered a “proper unit” by querying:

my_temp.unit

which would return the string 'K'. Or, if you need to remind yourself which unit did you use, you can retrieve the original value through:

temp, unit = my_temp.value

You can check the list of physical units available via:

my_temp.list_units()

The documentation for propertunits can be found here

What properunits does not do

Basically anything else. I wanted to have something that would allow me to connect the experimental side of my research with the models and simulations I create. I just needed to solve the IO problem, not to be able to work with arbitrary units. I may entertain in the future the idea of being able to use other units as the “proper unit”, but I must admit it is fairly low in my priority list. What I may do is expand the range of units to include specific use cases, like nautical miles and the like.

Webinar for physics of plasmas

After more than 20 years this year I published a new paper on the journal Physics of Plasmas where I explored how we can leverage machine learning to optimize processes in plasma enhanced chemical vapor deposition. Two relevant things about this paper:

• It is published with an open access license so anyone can access it (in fact, the whole Physics of Plasmas is open access for 2025!).
• It was one of the featured articles by the journal, and I was asked to give a webinar this fall about the paper, where I provided an introduction to plasma enhanced atomic layer deposition and our machine learning work. The link to the webinar is here.

Our review on Nature Reviews Methods Primer was published

This is a work that was a few years in the making. I had the chance to collaborate with some of the most talented researchers in my field to write an introduction to atomic layer deposition. The link to the paper is here. It is, alas, behind a paywall.

Chairing a session on quantum computing at SC2025

SC2025 is a conference focused on high performance computing. However, it has a post-Moore computing component, and as a member of the program committee for 2025 I was asked to chair one of the sessions in quantum computing

Others

• Honored to be a member of a PhD committee, Anurag did an outstanding job defending his thesis
• One of the projects I am leading got renewed this year, so more fun on ALD + microelectronics.
• Lots of changes in DOE, still processing them.
• Going through a lot of proposal reviews right now.

Overall, a busy couple of months.

Mis dos sitios en español son:

• anglyan.net

• cienciaypython: tutoriales usando jupyter book y MyST.