| HN Mirror

Specifically chemical process simulators used by Chemical Engineers are all very, very long in the tooth. I sold and trained other engineers to use these products for years. They are fine in the sense that thousands of engineers around the world use them on a daily basis to get work done successfully, but they could be so much better. I'll give you one specific example. Engineers will spend hours, days, and sometimes weeks (I'm not joking) trying to get simulations to converge without any luck because they fail to manually generate 'good' initial guesses for the models. And because engineering design is an iterative process, you don't need to just generate good guesses once, you'll need to do it many times as you iterate towards a final design or solution. A neural network could be trained and used to assist generating these initial guesses. I built a prototype in Tensorflow that did exactly this and the results were very good - of course this was trained on a dataset generated for a very specific plant. But it worked, and it worked well. But all of the companies in this space are stuck in the early 2000s and just want to keep doing what they've always done, there is very little innovation going on.