About
Three chemical engineers who wanted the model to be right.
Twin Process AI was started by three engineers from IIT Madras with backgrounds in chemical engineering and machine learning. The order of those two matters, and it shows up in every decision the product makes.
Why fired heaters first
Because a platform claim is easy and a working model is not. A fired heater has everything the hard version of this problem needs: combustion chemistry, radiant and convective heat transfer, a fouling process that nobody measures, real safety limits, and a genuine operating decision at the end of it worth money. If the method works there, it transfers. If we had started broad, we would have something shallow in six domains and nothing anyone could check.
It is also the largest single fuel consumer in most refineries, which means the arithmetic works out even when the improvement is small.
How we work
The physics comes first. The twin is an engineering rating pass: Shomate enthalpies, a Lobo-Evans radiant section, a rated convection bank, an ASME PTC 4 loss ledger. Machine learning sits on top to make it fast and to infer what the instruments miss. Not the other way round.
We publish what failed. A soft sensor that scored 0.05 is named on this site and switched off in the code. The gap between our optimistic accuracy figures and our honest ones is on the technology page. This is not modesty, it is the only way anyone can tell the difference between a model that works and a model that has been fitted until it looked like it did.
We do not claim plant results we do not have. Every number on this site comes from a first-principles simulation. Calibration against a live heater is our next milestone.
Where we actually are.
Pre-pilot. The simulator, the soft sensors, the optimiser and the validation are built and tested. What we do not have is a heater in a real refinery to calibrate against, and that is the next thing we need.
If you run one, the trade is straightforward: you get the analysis done properly and at no cost, and we get a model that has met a plant. We will tell you where it disagrees with your unit rather than quietly tuning until it agrees.