Welcome, everyone. This is Vinod Manamala, and I'm a senior data scientist. Today, I'll show you like how AI powered digital twins can transform the manufacturing operations. I'll also share the practical solutions from real factories that I saved like millions of dollars. Now, let's start by looking at what we'll cover today. So here's our plan for today. we'll start with why digital twins often fail in factories. Then I'll share five patterns that actually work, supported by a real, use case study. And, we'll also end with the practical tips you can use immediately. So first let's, examine why so many digital twin projects don't deliver the expected value. most digital twins fail in factories for simple reasons, right? so they can't handle the messy reality of manufacturing. That's the first thing. And they don't adapt to changing conditions. And they also ignore, sensor uncertainty. And, they try to model everything at once. And, and they also break when, the network fails. So these problems cost like real money when production stops or the quality drops. Now let's look at the first winning pattern that actually solved many of these challenges. The first winning pattern is what I call the even first architecture. So instead of tracking, fixed dates, here we track, every event that happens and calculate the states as needed. So this records complete history and helps you trace problems back to their source. In a chip factory, this solved, major headaches with tracking wafers through rework processes and saved, countless hours of troubleshooting. And, on the next slide, I will examine how AI can be taught to understand time in manufacturing processes. The second pattern is AI that truly understands time. In manufacturing, different time periods matter depending on what you're making. Our system uses special neural networks that learn which time periods matter most for each process. This improved prediction accuracy by 27 percent and spotted problems hours or days earlier, giving operators time to fix issues before causing, scrap. So this is really important. And now, next, let's look at how we can make these models adapt automatically to changing conditions. So the third pattern here, which is basically the models that adapt automatically. Actually solves a big problem with AI in factories. Knowing when to learn and when to hold steady. That is the, that is a big problem here. Our solution adjusts learning rates automatically. During stable production, it increases the learning to catch these subtle changes. And during unstable periods, it increases learning to avoid overreactions. So this reduced the false alarm by 62 percent and kept the models accurate by 3. 4 times longer without, manual retraining. Now, let's talk about how to deal with another major challenge. basically it's the uncertain sensor data. The fourth pattern here addresses the sensor reliability. The factory sensors get dirty, drift out of calibration or fail. And our approach here is very honest about uncertainty. The system here calculates the, confidence levels for every prediction, and tells the operators when, data looks suspicious. When the data is questionable, it falls back on physics based rules, and this cuts the false alarm by 45 percent while catching the actual problem. Now, let's look into the, fifth pattern, which addresses how manufacturing happens across multiple timeframes simultaneously. The fifth pattern here, recognizes the manufacturing that happens at many timescales at once. Our solution here tracks processes at multiple timescales, like simultaneously, and, this found problems that would be invisible otherwise, right? in one factory, we discovered morning temperature fluctuations were slowly degrading, a small catalyst over weeks. And, the system can connect a brief even from, weeks, that happened, weeks ago to a quality problem that's happening right now. this is pretty, amazing. And, now let's see, like, how these patterns work together in, in a real semiconductor factory, in a real use case. So let me show you how these patterns worked in a real semi, a real semiconductor factory. So their problem was handling the wafer rework. so when wafers need to repeat like certain process steps, and, they found this was like really problematic. So their first, digital twin actually lost track of wafers and got confused. So we implemented like an event based system with AI for like defect detection, anomaly detection and process relationship mapping. So this resulted in 94 percent accuracy predicting, reworks that were needed. And also like 12 percent yield improvement and also 4. 2 million in annual savings. Now let's take a closer look at how, AI actually worked in this factory. So here the, AI had two key components. First is the defect detection using the computer vision, which analyzed the wafer images with almost 99. 2 percent accuracy for, critical defects. Second, the process relationship, the process relationship mapping. Connected over, 1, 200 process parameters. So this revealed, cost effect relationships nobody knew existed. And, the engineers who were working there could ask, what would happen if you change this temperature and get realistic predictions before making, the actual changes. Now, let's talk about some practical implementation tips, starting with how to handle, all these data, these systems generate. The factories generate, enormous data, often terabytes daily, right? you can't keep it all, but you can't just delete it either too. our approach here is keep high frequency data for 24 hours. and save hourly summaries for months. The smart part is automatically, detecting and preserving the anomalies while compressing the, the normal data when nothing happened, generally. And this reduces the storage needs by almost, by 78%, while keeping 95 percent of the analytical value from those data. So the next implementation challenge we need to address is the, the network reliability in these factory environments. And, let's see, like, how we can handle these, network problems. The factory networks, they fail constantly due to, equipment interference and maintenance. and maintenance is the biggest problem, in, day to day activities. our solution here, what we did was, It puts the intelligence at the edge, next to the equipment. So the edge devices are like, they run lightweight AI models, and they work independently during the outages or, during the maintenance. So when the network reruns or when they come back online, they automatically sync with the, the central system. So this happened in a chemical plant and we achieved like almost 99. 97 percent uptime despite all these frequent network issues. Be it like, the equipment interference or a maintenance process. Now let's also discuss how to keep the AI models performing well over the time. So basically here, We're going to talk about the ML model lifecycle management. So the AI models in the factories degrade over the time as processes change. So our system continuously monitors model performance, and also it detects drifts in the models automatically. So the new models are first tested in something we call it as the shadow mode, before gradually taking control. So this reduced the, by doing this we reduce the engineering time. on model maintenance by almost 67%. And extended the model life almost by 3. 4 times with almost consistent performance across all the product changes. Now let's look at how the physical loss can make those AI predictions like more reliable. the manufacturing must follow like, physical loss. You can't create or destroy mass or energy, right? So here our system combines like AI flexibility with physical constraints. So this ensures the predictions obey the real world rules, even in conditions never seen before. In a chemical reactor, for example, this approach reduced the impossible predictions by 94 percent 94 percent and improved the performance in new situations by 72%. making this operation really safe. So now let me wrap up with key takeaways from, today's presentation. to summarize, the key takeaways here, the basically use even based architecture for complex manufacturing. and the second key takeaway is like implement AI that understands time and uncertainty. Third key takeaway is like build systems that adapt automatically, right? And the fourth key takeaway is like design for real world challenges like network failures that we spoke about. And the fifth key takeaway is measure success by business impact and not by technical metrics. And also start with a focus problem and then you can expand on it. So these are my key takeaways and thank you for your attention. And I'm happy to take questions. on my, using my email ID, you can connect with, also you can connect with me on my LinkedIn profile that you can see in the bottom. so I'm happy, we met on this Conf, 42 conference and I'm hoping to catch up with you guys. Thank you.