AI Models in Automation

I’ve been reading about AI since Sydney (a shadow self of the Bing generative AI) told a New York Times Tech Reporter to leave his wife to be with it instead [1]. I’ve been trying to find ways to use Chat-GPT, DALL-E, and other generative AI to make my life easier with little real success. But I’ve also read about much smarter people doing some really interesting things. And I am hot and cold on AI.

I am cold on AI (Large Language Models)

I am cold on Large Language Models – chatbots like ChatGPT, Copilot, Gemini, and LLaMa. While I do acknowledge that some jobs will move from writing to interfacing with a LLM I don’t see any lights flashing in our industry. You could get the model to write solve a Python problem to get a 9/10 on your professor’s homework set but I wouldn’t deploy anything written by a LLM in a manufacturing environment without so putting so much work into validating it that I might as well of written it myself.

Maybe I’m just getting old but I couldn’t even get GPT-4 to write an interesting blog post for Automation World (or even a passable one that would pass a Turing test smell check).

I am hot on AI (Convolutional Neural Networks & Monte Carlo Tree Searches)

If you have a chance read Hamid Khodabandehlou et al’s article on using Convolutional Neural Networks (CNNs) with Raman spectroscopy [2]. Their paper looks at the use of CNNs to create an accurate, generic model of a process to predict attributes beyond training data. Conventional methods (Partial Least Squares) limit the use of Raman spectroscopy to the quality aspects that it is based on.  Similarly, Monte Carlo Tree Searches, the same form of AI used to beat top ranked GO players, has been used to discover a new class of compounds to kill drug-resistant bacteria [3].

While these forms of AI provide important advances for humanity, the reason I am excited about them for automation is that they are far more results driven than the content driven (and on occasion hallucinating) LLMs. At their most basic, results are testable and verifiable. But beyond that I see clear applications within Process Automation. The nearest term, minimum viable product is probably tuning PID loops. Beyond that, Khodabandehlou’s paper shows strong promise to accelerate process development times by accelerating the development of calibration curves for process parameters. Additionally Neural-Network Based controls have shown promise for dealing with highly non-linear systems that are highly flexible and less noise prone [4]. The application of these forms of AI may still be a way off in our industry, but they could have a far greater impact than LLMs.

William TomHon is a Sales Director for the Process Automation division at Catalyx North America.

[1] https://www.nytimes.com/2023/02/16/technology/bing-chatbot-microsoft-chatgpt.html

[2] https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/10.1002/bit.28646

[3] https://dspace.mit.edu/handle/1721.1/153216 [4] https://www.mdpi.com/2075-1729/11/6/557