Transcript
Ravi: Welcome back to the Smart Factory Podcast series. My name is Ravi Jaikumar. I’m the Global Product Manager for Smart Factory Scheduling ATP Solutions, and today, we are switching seats. For the first two episodes, Sam asked the questions and I did the talking. Today, I’m asking because the topic isn’t really mine. I’m joined by Sam Duchscherer, who leads the AI products and solutions for the APG team. Sam, welcome to the other chair.
Sam: Thanks, Ravi. It is a little strange being on this side of the mic, but I am happy to be here with you and do this for the first time.
Ravi: Okay, I’m excited. So, I’m not going to ask whether AI works in a fab. I think that’s a boring question. I want to ask which things. And I’m going to ask the person on our team with every professional incentive to oversell it, Sam, so no pressure. One ground rule we agreed on before we started. We each have to name something our own side gets wrong. I’ll say mine later in the episode. So let’s get started.
Every vendor at every conference is saying AI now. Some of them are saying it about products that I knew were like rule-based or heuristics 18 months ago. So what’s actually different this time? Because manufacturing has had AI waves before, and most of them did not survive contact with the fab.
Sam: It’s a fair question, Ravi. And I’ll start by saying it’s okay to be suspicious, just like the last two podcasts. I mean, everyone is well aware that manufacturing has done this before. There were expert systems in the ’80s, neural nets in the ’90s, and a big data phase where a lot of fabs built a very expensive data lake and ended up with a very expensive swamp. So, when answering what is actually different, maybe I try to focus on three things. First being the data finally exists. Most fabs now have years of event-level MES history: every move, every hold, every tool state change. Twenty years ago, the models had nothing. I won’t say nothing, but you know what I mean, right?
The second thing that has changed is compute got cheap, right? Cheap enough to learn from history and simulate decisions in minutes instead of overnight. And lastly, this is the one people skip: The methods just got better at decisions, not just predictions. The old ways were mostly predict a number. Now, now we can learn a policy. Policy meaning what to do next, given everything that’s going on. But what hasn’t changed is the marketing, But what happens to change is the marketing. If something was a rule engine 18 months ago, and it’s AI powered today, it may just be the same rule engine in a new outfit. So, here’s what people could do if they’re skeptical. Ask what the system learned, from what data, and what it does differently. And if no one can answer that, then yeah, it’s an AI sticker, not a new capability.
Ravi: Okay, systems that learn. Learned policy. Why should a learned policy be the well-tuned heuristic? And pitch it to me specifically. I’m the person who just spent two episodes talking about queuing theory and telling people to go fix their availability. Make the strongest case you can.
Sam: Okay, let me do my best pitching this to the queuing theory guy here. First, dispatch isn’t one decision. It’s thousands of decisions a day, in sequence, where every choice changes the choices you’ll have later—and you make them without knowing what’s about to break or arrive. That’s essentially sequential decision making under uncertainty. It’s pretty much the textbook shape of the problem reinforcement learning was built for. And that’s not me being a fan or cheerleader of reinforcement learning. That’s just the shape of the problem.
Now look at what a heuristic is. A dispatch rule is a frozen trade-off. At some point, someone decides how much a due date matters versus a setup, versus a queue time risk, and baked it into weights. And to be fair, it was probably a great decision on the day they made it. But the right trade-off moves. It moves with product mix, with WIP level, with which tools are up. For example, at low WIP, maybe you chase due dates. But at high WIP, with a litho tool down, the smart move here is completely different. The rule doesn’t know that the world changed. It keeps doing what it was told—very confidently, I might add. It’s kind of like my GPS telling me to turn left into a lake.
Second argument I have, it’s around anticipation. A heuristic reacts to the current state, and a learned policy can act on the predicted state. So, for example, if I’m fairly sure a tool goes down in two hours, I should be making decisions that a rule can’t make because the rule only sees now.
Third pitch I have to you is the look ahead. Batching, setups, and queue times all interact. Do I hold the furnace to build a full batch knowing the lots I’m waiting for have already started their upstream queue time clock? That’s a really hard question. And you can write a rule for it, and people do, but you end up with a rule that reads like a tax code. So, what really wins? Well, I guess I would say quick, multi-objective trade-offs when nothing is stable. Overall, a well-tuned heuristic is a really good answer, but it’s a really good answer to yesterday’s fab. A learned policy, however, I would argue, is trying to answer today’s fab every 15 or every 30 minutes.
Ravi: Okay, that makes sense, but let me continue probing. If you recall, just a call back to the first episode, my argument there was that the real lever is in utilization, its variability, and halving your variability buys you more cycle time than any dispatch rule I could hand you. So, is prediction just variability reduction by another route? You can’t stop a tool breaking, but if you can see it coming, it stops being variability and becomes information. Or is that too generous for my own framing?
Sam: No, I don’t think it’s too generous. I think it’s the best version of my argument, and you made it for me. So yeah, variability hurts because it’s a surprise. You know, a tool going down isn’t really the problem, but a tool going down that nobody planned for is the problem. If you see it coming, you move work, you re-sequence, pre-stage the batch. The failure still happens. It just stops being a surprise. So again, prediction turns variability from a cost into an input.
But here’s where that breaks, because it does break. First, prediction doesn’t remove variability. It just lets you plan around it. So, if your tools are genuinely unstable, a great forecast just tells you very precisely that your day is going to be bad. You still have to go fix the tool. Another point where it breaks is a prediction is only worth something if there’s a decision that can use it. So knowing a tool goes down in two hours is great if you have a qualified alternative. If you don’t, well, congratulations, you just built a very accurate anxiety machine.
And then lastly, where it breaks is: predictions are just sometimes wrong. If the model says a tool is going down, you reroute everything, and the tool doesn’t go down, you just created the chaos you were trying to avoid. So, my answer is that prediction is variability reduction, but by other means. And it really is that only when it’s accurate enough, early enough, and maybe attached to a real decision. If you miss any of those three, you know, we’re right back to where we started on your previous episodes, Ravi.
Ravi: Okay, so you came well prepared for the questions. So now let me get to the question that I really wanted to ask you. I want you to tell me as the AI product manager, where should I keep my traditional systems? It could be heuristics, could be optimization, and not let you anywhere near it. Where does it not win?
Sam: Oh, Ravi, this is fun. Okay, so let me put on my least favorite hat. And the first example maybe is easy, right? Anything that’s a hard rule. For example, queue time limits, tool qualifications, reticle availability. Those aren’t really preferences. I don’t want a model learning its way toward “mostly compliant,” because in a fab, mostly compliant is a fancy word for scrap. Second area, I would say is where the problem is just simple. A stable tool set, maybe low mix, not much interaction upstream or downstream. A good rule is probably already close to optimal in these cases. Putting a learned policy on it is just like hiring a chess grandmaster to play tic-tac-toe. Sure, they’ll win, but you were already winning.
So, I would say that’s a pretty good list.
Ravi: Okay, thank you for that answer. I think for me and for our listeners, that question is a crucial question, where to use AI and where not to. Let me throw you another call back. In the last episode, I said something regarding the metrics, right? So, whatever you make the single metric, it will eventually get optimized in ways that you didn’t intend. And that’s not because it’s a fab thing, it’s because it’s the human and the people thing. In that case, I was talking about moves. But isn’t the reward function exactly the same problem, except worse? At least moves was printed on a shift report where everybody could argue with it. But a reward function is a KPI choice buried in a training pipeline that nobody outside the team ever sees.
Sam: You make a really good point. You know, a reward function is a KPI. It’s just a KPI written in code, maybe tuned by a data scientist and handed to a system that will do anything—and I mean, anything—to make that number go up. If the reward moves, you get moves. The easy lots, the convenient batches, everything from your list on the last episodes. Difference is, the model will find versions you haven’t thought of yet. It’s really creative. I would describe it maybe like a toddler who found a cookie jar. However, here’s where my side gets it wrong. We treat the reward function like a modeling detail, but it’s actually the most important decision in the whole system. And it often gets made by the person furthest from the business. So, what do you do about that? Well, maybe some suggestions. First, you could write the reward down in plain language. This policy is rewarded for X, penalized for Y, and here’s the weighting. If leadership can’t read it and nod, then I would say it’s not ready. You could also look into holdout metrics. Pick metrics that are deliberately not in the reward, like stability, on-time delivery, or queue time compliance, and watch them. If the reward goes up and the holdouts go down, the model is probably gaming you. And lastly, it’s just always a best practice to do shadow comparison before it touches the floor. Run it alongside current dispatching, log what it would have done, and have humans look at the weird ones. The weird ones, the weird outputs, are where the reward function is telling you what it really means.
Ravi: Okay, I’m raising my hand for a follow-up on that. Who reviews the reward function? Because a KPI change goes through a governance process typically in a fab. Does a reward change?
Sam: Okay, so you’re right. Honestly, in most places, we’re still hardening this answer, and I’m including most in the industry today. So a KPI change at a fab goes through a meeting, a slide, someone in senior leadership asking, “Are we sure?” However, a reward change is often a code change. You know, maybe one engineer can move a weight from 0.3 to 0.5 on a Tuesday. The tests pass and the behavior on the floor changes. There does seem to be a governance gap. And I’d rather not lie on this podcast and pretend that’s 100% solved. Overall, what I can say that we are aiming for, or what it should look like, is that a reward change goes through the same review as a KPI change. Leadership in the room, not just a data scientist. It’s also versioned, so you can see what the policy was rewarded for in March, maybe versus today.
Ravi: Okay, now to the topic, when it comes to AI, the topic of trust and explainability is always on people’s minds. If your operators— this is something I mentioned in the previous episode—if the operators are overriding 40% of recommendations, then that means every other number we are reporting describes the system that isn’t actually running. So, imagine the situation, three in the morning, a shift supervisor is looking at your recommendations and asking, why lot X went before lot Y. What does your system say?
Sam: Well, maybe we start with what it should not say, right? If it says lot X has a priority score of 0.87, to me, that’s not an explanation. That’s a number wearing a lab coat. And nobody wants that at 3 a.m. They barely want that at 3 p.m. What it should say is one line in supervisor language, something like lot X first, its queue time window closes in 40 minutes; lot Y has six hours of slack. So, we list a reason, a number they can check, and a consequence—two seconds to read. Even better, if it tells them what happens if they override. If Y goes first, X misses a window. That’s what turns a black box into a colleague, in my opinion. Now, the honest part: for predictions and constraints, that’s very doable. For a full learned policy, it’s hard. The model doesn’t really have a reason the way a rule does. We’re reconstructing the reason after the fact. So, is it solved? No, but is it where I’d want any team to spend the most effort? Absolutely. The best decision nobody trusts is worth less than a decent decision everybody follows, in my opinion.
Ravi: I agree with that, 100%. Sam, should a model ever get a vote on feasibility, or does that layer stay deterministic forever? I’m referring to queue time links, qualification status, radical availability, things like that.
Sam: Okay, so I’m not sure if our listeners were expecting us to fight. Hopefully not, but I’m here to disappoint them if they were, because I agree with you, right? Completely. Feasibility stays deterministic. Q time, links, quals, radicals, the model doesn’t get a vote. And the reason is simple. A model that’s learning a constraint is by definition sometimes violating it while it learns. In a fab that’s not a learning experience, that’s scrap with a training log. Where AI can help is around the constraint layer: predicting when a queue time window is at risk, so you act earlier. If there’s one place an AI product manager should be most conservative person in the room, it’s probably that one.
Ravi: Okay, here’s my scheduler question, Sam. A policy that’s 3% better on average and replans the queue every 15 minutes, is that a win or a loss? Because people plan around the schedule. If it thrashes, they stop believing it. And then the 3% is just theoretical on a paper.
Sam: It would say loss, and maybe I walked right into that one. I mean, because what, 10 minutes ago, I said every 15 minutes, like it was a feature. You know, people plan around the schedule. Operator, stage, material, maintenance, plan, PMs, the next shift reads the plan at handover. If it reshuffles every 15 minutes, they could stop believing in it. And that 3% only lives in the spreadsheet. But here’s the engineering point. The fix isn’t to let the model thrash and then smooth the output afterwards. That’s a band-aid. Stability belongs in the objective itself. Changing the plan should have a cost, so the model only changes its mind when it’s worth it. Kind of like I would say a person — they’ll change their lunch plans for a good reason, not because there are slightly better sandwiches two blocks away. Put stability in the objective and the model learns to be boring on purpose. And in a fab, boring on purpose is a feature.
Ravi: Okay, I have two questions and they are related. First, like you know, Sam, in our APG, we work with a lot of customers who are building greenfield fabs. So greenfield, new fab, ramping, no history. What does a learn policy do when there’s nothing to learn from? And the second question is, this one’s again a callback. Last episode, I said a demand downturn makes any scheduler look brilliant because loading drops and then the queue time falls on its own. So, what happens to a model that was trained during a soft quarter when volumes come back?
Sam: I will take the second one first because it’s a little less scary to answer. A model trained in a soft quarter learned a world with short queues and lots of slack. When volume comes back, the world doesn’t exist anymore. The model is still very confident. It’s just confident about last year. It’s like learning to drive on an empty Sunday morning and then hitting rush hour. You won’t prevent that. You or automated methods just might have to detect it. You have to watch whether the inputs still look like what the model was trained on and watch whether its predictions are still landing. I would say prediction errors usually creep up before the fab KPIs move, so that’s your early warning. The goal is for the teams to spot this drift before the fab manager does, because once the fab manager notices it, you’re already late.
Now the greenfield question. with no history. Well, a learned policy with nothing to learn from isn’t a policy. I would say it’s a hope. So, in practice, you start with heuristics. Yes, heuristics. You pre-train in simulation carefully, knowing the simulation is going to be wrong in some places. You transfer from similar tool sets where the data allows, and then you stage for autonomy. First, it only predicts, then it recommends, and only with evidence does it act and act on a narrow slice. Honestly, the full greenfield ramp up story deserves its own episode, and probably with someone who’s actually lived through one. But I hope that answers your question.
Ravi: Yes, it does. And I’ll be looking forward to that greenfield episode anytime you do that. Now, onto the thing we agreed in the last episode. I said that it’s not very easy to AB test a fab. You can’t AB test a fab. It’s one fab, one set of WIP and everything coupled. But now the thing I’m validating changes over time. So how do I know it’s still good six months in? And more importantly, who owns retraining and on what trigger and with what gate before it goes live?
Sam: So, validation can’t be a one-time event, in my opinion. It’s less like a driver’s test and more like a pilot’s reoccurring check ride. So, things to look for: monitoring. You’re tracked continuously; it’s not a quarterly review. Also look for drift detection on the inputs and on prediction accuracy. Another thing is defining trigger for retraining, like drift passes a threshold; there’s a new product, a new tool set, or not retraining because someone felt like it. A gate is also important, meaning the retrained model has to match or beat the current one, then run in a shadow for a while, and then get sign off from someone on the operation side. Now the uncomfortable part, who owns the retraining? And that, honestly, I don’t have a good answer. We see different factories wanting different things. What I will say is that if it’s the responsibility of the vendor, if that responsibility ultimately sits with the vendor of retraining, we should capture the triggers, the expected timeline, and the agreement, so the fab isn’t left managing the operational uncertainty.
Ravi: Okay, that makes sense. Now, let’s land it. And I know this is not a binary, but if you drew the stack on a whiteboard, what learns and what doesn’t?
Sam: Okay, let’s grab a marker. Now, this is a little bit difficult because it’s a podcast. But consider four layers, you know, bottom to top. Bottom, we have constraints, deterministic, things like cue time, quals, reticles, recipes. This layer decides what’s allowed and nothing learned gets a vote. Now, the second layer is prediction—things like arrival, downtime, cycle time, when a batch is likely to fill; everything we said about turning variability into an input lives here. Third is maybe optimization. It takes the predictions, respects the constraints, balances the objectives, including stability, and produces a decision with a reason a supervisor can actually read. Now the fourth layer, I’d say the top layer, would be the human. The human keeps the override. It’s the most honest signal of whether everything underneath is working. So, at the bottom of this whiteboard that doesn’t exist, the question can stop being AI or rules. The real question is which layer of the stack should learn. And in a fab, that’s the prediction layer, not the constraint layer.
Ravi: Okay, I think our respective viewpoints are now coalescing because I do agree, heuristics don’t scale to the trade-off complexity we are now dealing with, especially with advanced packaging flows and hand-tuning the rule weights across 100 tool sets isn’t really a strategy. It’s a maintenance burden that decays quietly, and it decays fastest exactly when the fab is changing most, which is when you need it to be right. So, when I defend the well-tuned heuristic, the word doing most of the work is well-tuned. And in practice, most of them are not, not because anyone was careless, because nobody has the time. So let us look forward now, because this is where I actually want your opinion and not your position. Three years out, what does scheduling look like? And I don’t want the roadmap. I want your honest guess, including the part you’d be embarrassed to be wrong about.
Sam: So, I would say these are educated guesses, right? We have been working with quite a bit of customers at Applied on some of these, but let me give it a go. Prediction one, which is more safe-ish, I would say, is that most leading-edge fabs are using learned predictions like arrivals or down times and cycle time, and those will feed into dispatching decisions by default. So, I would say in this stage, we would go beyond pilots and separate dashboards. This would just be the new plumbing. People will stop calling it AI, maybe the same way that no one, you know, calls out spell check anymore.
Prediction two, that is, let’s say if I was embarrassed to get it wrong, is, you know, fully learned dispatch policies will be running in narrow places like specific batching areas or specific bottlenecks.
But prediction three, the biggest change, I would say, won’t be the model; it’ll be the inference. Supervisors will ask the scheduler questions in plain language. Why is this lot waiting? Or what happens if I pull this tool for PM now? and get answers in seconds. The decisions will look similar. The conversations around them might be completely different though. So, I guess to summarize, I would say start with plumbing and then maybe narrow autonomy and a scheduler, then, that you can actually talk to.
Ravi: Okay, now this is interesting. The next question that I want to ask. Everyone’s seeing agents now, and that includes me. I don’t know, in the last six months or 12 months, I’ve used that word so many times. What would an agentic scheduler even mean on a fab floor? And what would have to be true before anyone should let one act without a human in the loop?
Sam: Yeah, agents or agentic is for sure the word of the year. Here’s the real distinction to me. Today, most scheduling systems recommend into one system; run this lot next. An agent in the real sense uses tools, plans multiple steps, and acts across systems. So instead of, “Run lot X,” it notices a queue time risk, checks the alternate tool qualification, checks where the reticle is and tells the supervisors what it did and why. To me, that’s the real capability change, not a rename. But before anyone lets that act without a human, a lot has to be true, in my mind. The constraint layer has to be airtight because the agent acts right across it. The data has to be trustworthy, because let’s say if the retical status is wrong, an agent’s just going to act on the wrong world faster. Also, every action has to be logged. reversible, and limited in what it’s allowed to touch. It has to explain itself before it acts, not after. And it needs to track record, months of recommending with a low override rate before it earns a narrow lane on its own. So, is the industry there yet for a hands-off agent on the fab floor? Maybe not at the full fab level, but agents that do leg work and let humans click approve? Yeah, yeah, absolutely.
Ravi: Okay, I think we are coming to the end of the podcast, but I have a question that is perhaps the most useful one for anyone listening. Sam, what should a fab be doing right now to be ready? And I mean doing, not buying.
Sam: Yeah, good distinction. First, I would say clean event history. So, every move, hold, and tool state change with a timestamp you trust. Gaps in your MES history gets inherited by every model you’ll ever build. Simple one is maybe label your downtime. I would say start measuring your override rate today, even with zero AI anywhere. So how often do people override your current dispatch system and why? That’s your trust baseline, and you’ll want it later. Four, how accurate are your quals, your queue time limits, your reticle status? Not on paper, on the floor at 3 a.m. And I would say a bonus is start writing down your KPIs, what they actually are and who owns them. Because one day someone’s going to turn them into a reward function. and you’ll want to know you picked the right one. If you do all of this, you’ll get better results from whatever you run.
Ravi: Super. Now, it’s time for the game. I think we concluded the podcast, but we cannot end the podcast before the myth versus reality game. So, Sam, you put me through this twice, so it would be rude to not return the favor. I have two myth versus reality questions. The first one, you need a mature digital twin before AI can do anything useful in a fab. Is it a myth or reality?
Sam: I’d say mostly myth. It depends on which layer you mean. For prediction like downtime, arrivals, and cycle time, you don’t need a digital twin. That trains on historical event data that’s already sitting in your MES. You just need clean history, not a digital twin. For policy learning, where the model has to try decisions and see what happens, you can’t experiment on a live fab. You will need a high-fidelity environment to learn in, and the twin really is a prerequisite. The trouble is when people mash those two together. We’ll start AI once the twin is ready. Well, the digital twin takes time, so don’t make the easy win wait for the hard one, if that makes sense.
Ravi: Yes, it does. Now to the second question. Eventually, AI replaces the dispatcher and/or scheduler entirely. Myth or reality?
Sam: I don’t see full autonomy coming for the dispatching role, not in a quick timeframe that matters to people listening. What I would say, what changes is the job. Today, a dispatcher makes a lot of individual calls. Over time, the system takes more of the routine ones, and the human moves up, so if anything, the best dispatchers become more valuable, not less. So, I would say for the most part, that’s a myth.
Ravi: Great. Sam, thank you so much and for taking the harder chair because I’ve been asking the easy questions all episode. This was fun.
Sam: Thanks, Ravi, and these are definitely not easy questions. So, until next time, listeners, let’s keep learning together, keep questioning assumptions, and keep looking beyond metrics everyone else is celebrating. Thanks, Ravi. This was fun!
