Comment by wanderingmind | Hacker News Reader

Comment by wanderingmind | Hacker News Reader

High throughput material discovery is one of the biggest fad in academia that has been draining money for decades. They started with high throughput experiments, then combined it with physics based modelling and now ML.

However not a single industrial relevant material has been created in decades but hundreds of millions of dollars have been wasted.

The trouble is most of them cannot be manufactured at scale because they are some crazy metastable state that can only be achieved at lab scale. Further, when manufacturing at scale, materials will have impurities, defects and the properties of such exotic complex oxides are highly sensitive to such changes.

As a result hundreds to thousands of papers are published every year and nothing happens except the professors getting rich and graduate students graduating either unemployed or going to become software engineers. Another reason to remove the incentives to create PhD from slave labour.

I worked on projects in this area and honestly this isn’t far off the mark. Industry is desperate to get new materials to market faster, but after seeing the entire development chain, the scale-up is a far greater bottleneck than the generation of promising leads. You can even see some of it buried at the end of this paper - they’re using some optical property as a surrogate for catalytic activity, acknowledging that

> Catalyst testing for water oxidation is far more expensive than the optical screening from our high throughput workflow, and even though there is no known connection between the optical properties and the catalytic properties, we use the analysis of optical properties to select a small number of compositions for catalyst testing

I’m not saying these high throughput screening techniques don’t have value - they do and they have their place. They get overhyped because software companies see a potential ML application and throw their weight behind it. However the ML-based search algorithm is usually the lowest value-add to the process. Automating the characterization (the part the Caltech people did.. built a custom microscope) is the biggest de-bottleneck in this workflow.

So what is your suggestion?

Stop doing research in this area?

Which money gets burned? Yours? Or do you think that this money would have been allocated differently/better?

Ofcourse it's my tax dollars that's getting burnt through this useless endeavour. Academic funding must be for bold ideas that are more likely to fail but push the knowledge boundaries further. There is nothing being learnt by mindless number crunching that is done in such work. Infact this is worse than buying lottery tickets hoping to hit a jackpot.

Reply on news.ycombinator.com