Why physics-based AI outperforms theory (and vanilla AI)
We have all experienced it. There are beautiful theories that explain exactly how your product should work. And then there is reality.
Complex interactions between components, electronics, and real world environments can make the theories in textbooks hard to apply to your system. For some applications, the theory will get you far enough to model your system’s behavior, but for others, it does not come close.
My first experience training AI models was to help exactly this kind of issue. I worked for a company selling very precise, state of the art spectroscopy devices. They could measure H2S with tens of parts per billion precision. They achieved this by ignoring all of the science behind spectroscopy in their algorithms. Between the electronics, the laser, the detector, the mirrors, the windows, and the lenses, the absorption signals did not behave well enough to follow the theory. This meant that every system had to undergo a thorough calibration at the end of the manufacturing process. After every component was assembled and every complex interaction was included in the signal, they gathered a spectrum for each gas present in the stream at every pressure the application required. This process could take days, but the calibration gave them high fidelity and a competitive edge in a tough market.
The AI approach we took was simple. We wanted to know if an AI model could take one or more of those calibration spectra for a single gas at a single pressure and predict how the spectra would look at other pressures. In effect, we were trying to make AI solve the calibration for each device. This company had over a decade of data, and spectra are data rich, so the model worked fairly well and led to a patent (US11754539).
But the performance could have been better. The model was only truly successful working with one gas at a time, and in the pressure ranges they typically operated in. For instance, if you gave it a Methane spectrum at 900 mbar, it would very accurately predict the Methane spectrum at 1200 or even 1700 mbar. But why could it not predict what an Ethane spectrum would look like at 900 mbar? Or what if they got a new customer who wanted a device to operate at higher pressures, like 4000 mbar? That is well outside the maximum pressure in the historical data set, so the model likely would fail. The physics is the same, and the instrument is the same, so why could the model not do better?
Because back then, I did not know about physics-informed machine learning. The way I trained it, the AI model had to learn both the pressure broadening physics equations and the instrument function. It was able to do this with some success, but the performance is always going to be limited. Limited by your data, limited by your company’s history, and limited by the complexities of the real world. I do not work at that company anymore, but if I did, I would restart the project and bake in the well-established pressure broadening equations and have the AI learn the instrument function alone. This would require less data, be more extensible to new applications, and perform much better.
And in this case, a higher performing model does not just mean saving a few minutes or a few hours. A model that can use one gas at one pressure to calibrate a whole system opens up the possibility of field calibrations with a cheap bottle of methane. The customer could have a continually updated, high performing system that suffers from zero drift over time with a more flexible, extensible, higher performing AI model.
I may not be able to go back and fix my old, big data newbie beginnings, but now I have started Collimated Research LLC. We focus on building AI models with a grounding in physics, so you can make novel discoveries, achieve high performance in complex applications, and extend your product lines to new areas. Most importantly, we will help you model the nasty stuff that AI handles well (multi-objective, highly complex and particular systems) so the physics your system is based on can shine.
AI has the potential to push science and industry to a new level of discovery and growth. But why would you want to work with a model that is less technical than you are?
The team at Collimated Research have scientific training, and will pull in the latest academic findings in your field to build you models as smart and knowledgeable as you are, so you can push your innovation to new bounds.