With the surge in innovation in artificial intelligence (AI) thanks to progress in computing power, data gathering, and algorithm development, materials informatics (MI) for improving the workflow of chemical and material discovery has gained momentum as well. Many companies across the physical industry value chain have touted MI, and many of them (if not all) have been thinking about working with an MI startup. Despite their claims, it is often unclear what value these startups can bring to the table – and in particular, it's often hard to gauge whether working with a startup or developing solutions internally is the better route. In this blog, we dissect the pros and cons of each approach.
working with a startup
As we discussed in a previous report, working with a startup entails sharing data with it. While all MI startups claim no data reuse, they often keep the custom-developed algorithms for future use, thus still potentially giving future customers a "last-mover" advantage. Additionally, many MI startups have identified a major opportunity in building machine learning-friendly data infrastructure; companies like Citrine Informatics and Uncountable have even built their data platforms for smoother data collection and model training. However, these startups typically only provide the platform and some raw data (often a small portion compared to what you have to provide) and offer consulting services on how you can integrate your data from different sources on their platform. Moreover, most MI startups focus on optimizing the R&D workflow, not downstream production. As a result, the solutions they provide may not always translate to large-scale chemical production in a plant.
STARTING YOUR OWN TEAM
Regardless of corporate structure, all chemicals and materials companies have data scientists. One would assume it is relatively straightforward to start an MI team with data scientists/machine learning researchers, experimentalists, and computational chemists. However, there are knowledge and experience gaps among these groups, MI is a new discipline, and it is still very hard to find the right talent (we know a few research groups that may have good candidates). Furthermore, an internal MI team would need IT experts that can build data infrastructure and maintain its safety. Lastly, sometimes there exists a cultural resistance within corporate R&D with regard to MI deployment. The often-heard reason is machine learning's physics-agnostic nature – with enough data, the right algorithms can make good predictions without understanding the "why," but most traditionally trained researchers consider the approach "nonscientific," despite the potential increases in speed and new discovery.
While MI startups do provide good value on certain chemical and material problems, their value proposition on data and/or platforms is often not strong enough to justify a large-scale deployment. Instead, starting your own MI team should be the long-term strategy. That said, finding the right MI talent and changing major stakeholders' (not just R&D scientists') mindset is the immediate challenge, and outside partners can be invaluable in getting started and in overcoming some early cultural obstacles. However, companies should also consider a third way and work with academic groups on small-scale projects. Top researchers like Dr. Rampi Ramprasad, Dr. Taylor Sparks, Dr. Milad Abolhasani, Dr. Elizabeth A. Holm, and Dr. Klavs Jensen often have emerging solutions that have not been commercialized yet, which presents good partnership opportunities. Regardless of the specific approach you choose, do reach out to us to discuss strategies, targets, and partners.
- Executive Summary: How to Form a Materials Informatics Strategy
- Tech Page: Materials Informatics (for Lux members only)