2024 Reflections: Year of Learning and Building

As I look back on 2024, it's been a year filled with deep dives into AI/ML and scientific computing, with sprinkles of materials science, chemistry, and physics. I coverage for categories looks like this:

2024 category coverage

I figured it would be a good idea to write a summary of my posts for the year. I haven't done this in the past, but I think it may have some good value for me. Here are some of the highlights for the year.

Materials Science & AI Integration

The year started strong with my exploration of materials informatics tools. I worked extensively with the ASE Phonons class for phonon calculations, even creating a pickling utility to save calculation states. This theme of materials science tooling continued with my investigation of CrystaLLM, a fascinating LLM trained on crystallographic data.

A significant portion of my year was dedicated to implementing and understanding Graph Neural Networks (GNNs) for materials science. I documented my journey with Crystal Graph CNNs here, then here and provided updates on my progress at the end of May. The exploration somewhat culminated in working with MACE and LAMMPS for atomistic simulations, where I gave working solutions using Apptainer. This will remain an area of interest and focus for me.

Scientific Computing & Infrastructure

I spent considerable time improving my computational workflow and understanding. This included writing a post on GPU computing and implementing Lennard-Jones potential calculations to compare CPU vs GPU performance. I also explored tools for high-throughput computational chemistry, documenting how to democratize these capabilities. To date, this is one of my most popular posts that gets cross-linked from other sites, 🙌.

One interesting development was my work with electronic lab notebooks, particularly examining Prof. Kitchin's ELN approach using emacs org-mode. This tied into my broader interest in scientific documentation and reproducibility.

I documented my approach to using Materials Project API to streamline data retrieval for materials research, and finally got arround to packaging and releasing an old polycrystalline tool.

AI Safety and Innovation

I kept a close eye on AI developments, particularly around use in materials science and physics. I did have a post early in the year on safety and innovation, AI safety and Karpathy's work, and later explored exciting developments like Kolmogorov-Arnold Networks (KAN), which represent a fascinating new approach to neural network architecture. Interestingly, I haven't really followed up on KANs and what progress with them has been made, 🤔

Learning Methods and New Compute Paradigms

I formalized my learning approach in a post, breaking down how I tackle new subjects through priming, compressing, and authoring method. This methodology proved particularly useful as I explored new territory of thermodynamic computing here, then here, and most recent here. Still have a lot to learn on this topic and need to finish going through the book by Peliti and Pigolotti.

Other Insights

As my of my day job I had to start thinking a little about the challenges of simulation methods and computing setups. This lead me to spend some amount of my personal time looking into off-lattice, on-the-fly KMC simulations and the potential of remote disk mounting for efficient data management.

Looking Forward

As the year closes, I'm particularly excited about the convergence of materials science and AI. The rapid development and number of tools coming out in this space is super exciting, just hard to keep up and determine what the competing trade-offs are. The one area that I didn't write enough about on, was self-driving labs and still a lot that I need to read on and toy around with. I really think this is going to change the way materials development works is done in the future. My guess is that in 10 years is that decision making and operational tasks will be AI lead and orchestrated but top-level guidance will be done by humans.

So blog writing in 2024 has reinforced the importance of building strong foundations while staying current with cutting-edge developments. Whether it was understanding fundamental concepts like entropy or exploring new computational tools and methods, each investigation has added to my toolkit for tackling future challenges.

I would say the year has been marked by a balance between theoretical understanding and practical implementation, from getting around to posting on basic physics concepts like the Fermi energy¹ to implementing complex computational tools. This combination of theory and practice continues to be essential for meaningful progress in computational materials science. It was a good year of learning and decent in blog writing. Hoping to continue into 2025 and continue to push the boundaries of what I know and can do!

As this is my last post for 2024, I'll see you in 2025, and happy new year 🥳!

Footnotes

---

1. Turns out that I had already written a post on the Fermi energy, but forgot that I did so, and wrote a new post that ended up being similar but slightly different focus. ↩

Reuse and Attribution

Speeding up Commercial R&D Labs with AI

Post Reference Integrity Issue

The contents of this post were based off a reference to a preprint and that preprint has been found by MIT to be misleading and potentially fraudulent. Therefore, this post is no longer accurate or trustworthy with respect to reference [1]. I've decided to leave the post up for integrity and transparency, but much of my own opinions that were formulated from the reference are potentially incorrect and I do not recommend refering to or citing this post in any works. Here is the MIT release notice from the https://economics.mit.edu/news/assuring-accurate-research-record expressing lack of confidence in the work.

A recent preprint from an economics PhD student at MIT was published [1] on his homepage earlier this week. Its an interesting read because it assessed the impact of AI tools on a very large materials science commercial R&D lab¹. The major takeaways I took from the paper are:

• Increased Output: AI tool(s) led to a significant increase in the number of new materials discovered, patent filings, and product prototypes.
• Maintained Quality: New materials produced by AI tools meet the required standards without compromising quality.
• Enhanced Novelty: Claim that AI tool facilitated the creation of more novel materials, patents, and products compared to those developed without AI.
• Unequal Benefits: The benefits were not evenly distributed across scientists. Those with greater expertise and experience, particularly in evaluating potential materials, saw the most significant productivity gains.
• Shift in Tasks: Automation of the ideation phase allowed scientists to focus more on evaluating and refining material candidates.
• Importance of Expertise: The paper emphasizes the importance of domain knowledge, especially for assessing/refining AI-generated suggestions.

As usual with AI (hype or not), the impact and scope of the paper was discussing how AI will revolutionize the way researchers discover and develop new materials. The study highlights how AI did indeed boost such activity in a commercial R&D lab, but it was closely tied to the skills and expertise of the scientist and engineers leveraging the AI tools. The AI tools discussed assisted in the materials discovery process by allowing scientists to input desired properties with the model generating candidate compounds that are predicted to possess those properties. This interaction streamlines the research process and enhances productivity.

The papers reports that the introduction of the AI tool led to notable positive effects, including increased materials discovery, patent filings, and product prototypes. The study indicates that there was no evidence of quality compromises, but its not clearly what quality means here other than indicating that there was no sacrificing of standards. One important note, is that while the AI tool provides significant benefits, the advantages are not evenly distributed. More senior staff with expertise, particularly in materials evaluation and selection, saw the most substantial productivity gains. The most gained value comes from the perceived shift in research tasks to AI tools. This allows scientists to focus on refining candidate options rather than searching, but underscores the need for the users domain knowledge in the evaluation process.

Implications for the Future of R&D

In my opinion, theres no doubt ML/AI will be mainstays in scientific R&D, but I'm still uncertain how good current ML/AI tools are at accelerating scientific discovery, but it will undoubtedly get better. Maybe one exception is AlphaFold 3² as its success is unparalleled. I do see the tremendous value in the initial ideation phase, because many times you have a inkling of what your thinking in terms of design, but to actuate on it is often difficult because of lacking specific knowledge for a physical (i.e., material/chemical) system. With generative AI you can now more easily access such knowledge. I'm excited to see how this will all evolve as indicated by my regular post on this topic.

Probably the best visual for the impact can be seen in figure 5 (below) from the paper, which showcases how promising or impactful AI can potentially be.

Adapted from ref. [1]

In Panel A, the impact of AI tool(s) (labeled as "treatment effect" on y-axis) at the end of the study period: a 44.1% increase in new materials discoveries, 39.4% more patent filings, and 17.2% growth in product prototypes. It basically shows how the AI tools used in the commercial R&D lab fast-tracked each stage of innovation over a long period.

With Panel B we see more details in how the value is added over time. The effect on materials discovery and patents kicks in after about eight months, while product development takes a over a year to hit full stride. The data points prior to adoption of AI show where the baseline for the efforts were. Its a pretty impressive illustration of the gradual yet the impact of AI on innovation.

Its an interesting read and preprint to mull over as AI continues to enter our everyday research activities. The paper is pretty long and I didn't include all the research methodology specifics but its all there in the paper.

Footnotes

---

1. The paper does not specify the company only that it was a large R&D lab. ↩

2. AlphaFold 3 is a deep learning-based protein structure prediction tool that is poised to revolutionize the field of biochemistry. The lead authors of the AlphaFold paper were awarded the 2024 Nobel Prize in Chemistry for their work and it was recently released as an open source tool for non-commercial use. ↩

References

[1] A. Toner-Rodgers, Artificial Intelligence, Scientific Discovery, and Product Innovation, (2024). URL. [REMOVAL FROM arXiv REQUESTED]

Reuse and Attribution