DAIM: Inside The Algorithm | AI in Industrial Manufacturing: Inventory Optimisation and Real-Time Process Control

0:00The funny thing is that we were calling it the M-ventry planner, but in this case, actually, it's quite pertinent because we observed that this ensemble of machine learning and mathematical components sometimes was finding solutions which were sometimes similar to what the real planner with available tools was doing, but sometimes it was quite different. Sometimes this was reducing the frequency, sometimes was increasing the frequency, sometimes was timing the orders in a specific way. But it was like a non-trivial strategy, non-trivial order placing policy that this AI solution that ensemble all of these components was actually proposing.

0:39Dr Lucy Morgan:Welcome to Inside the Algorithm, the show that goes beneath the surface of artificial intelligence into the scientific research, technical breakthroughs and academic thinking shaping where this technology is actually heading. I'm Jeremy Bradley, Chief AI Officer at Canberra Spark, a leader in transformational data and AI upskilling, career development and progression. Each episode, I sit down with the researchers, data scientists and technical experts working at the true frontier of this field to explore not just what's happening in AI, but also the real science and thinking driving it forward.

1:11Dr Lucy Morgan:Whether you're building sophisticated AI systems, leading teams through AI transformation, or driven to understand what's really happening beneath the headlines, this is the show for you. In today's episode, we're going inside the data science operation of one of the world's largest consumer healthcare companies, and into two of the most technically ambitious AI projects it has produced. My guest is Dr. Georgi Mihailov, Principal Data Scientist and AI Director at Halion, also Visiting Research Fellow at King's College London and a mathematician by training whose PhD in pure mathematics from the University of Milan now underpins some of the most sophisticated AI systems running in industrial manufacturing today.

1:53Dr Lucy Morgan:Georgie led the development of Helion's AI Inventory Planner, a system that defines optimal corridors for inventory management and is now scaling into an enterprise-wide integrated business planning initiative. He's also been at the centre of the golden batch project and attempt to use AI for real-time control in pharmaceutical manufacturing, something that, as you'll hear, remains genuinely novel territory. This is a conversation about where the hard mathematics actually lives inside a global consumer health business, what it takes to move AI from a planning tool to a live control system, and what the end game looks like when you're trying to build supply chain excellence at scale.

2:34Dr Lucy Morgan:I do hope you enjoy it. Jogi, welcome to the podcast. Thank you very much for joining. Thank you for the invitation.

2:48Dr Lucy Morgan:So, Halion, for those I'm sure who are familiar, is a huge FTSE 100 business with very familiar brands like Sensodyne, Panadol, Centrum, brands that billions of people buy, interact with, certainly have seen in shops. Can you give us a sense of what data science challenges actually look like inside a business like Helion? Yes, the role of data science is quite relevant across all business domains. In fact, we have a relatively big data science team and we are supporting through various enterprise initiatives all business domains, like, for example, commercial, also marketing, R &D. We have a series of generative AI initiatives and a lot of important initiatives in my space, which is quality supply chain and finance.

3:48And yeah, there is a lot of demand and a lot of interesting and non-trivial mathematics trying to solve interesting and non-trivial problems, such as optimizing commercial investment, optimizing forecasting and optimizing marketing investment by following trying to anticipate consumer behaviors and in the space of supply chain obviously large-scale optimization problems in in logistics manufacturing planning delivery and so on and so forth and we are supporting across all these various domains.

4:31Dr Lucy Morgan:Yeah, that's fantastic. And it will get onto some of the supply chain challenges shortly because I know that's an area you've worked very heavily in with Helion.

4:48Dr Lucy Morgan:So let's talk about the inventory planning and in particular one of the tools that you and your team have worked on over the last few years, the AI inventory planner for Halion. So before we get into maybe how it works in detail, can you set the scene for us on what it's actually trying to solve in this space? What was breaking and not working well enough maybe and how the inventory planner managed to solve this problem? So inventory management is a centuries-old problem. However, recent technological developments allow us to have much, much better digital visibility on processes in complex supply chains such as ours.

5:37So basically, the purpose of the Aviation Planner was to leverage better the digital visibility that we have both upstream on supply. So how exactly our purchase orders are processed and then transported and then delivered. and also leverage better digital visibility downstream, having better demand forecasting and demand sensing capabilities and combining them in one ensemble of tools or machine learning models, actually machine learning models, which are able to improve best practices, improve, highlight further opportunities for improving inventory management across the network. So, yes, inventory management is quite important because if you're able to guarantee your service levels with smaller and smaller inventory in your network, this basically allows you a competitive advantage and allows release of working capital, which is always very important for companies.

6:39Dr Lucy Morgan:I mean, that's fascinating and beautifully encapsulates what is, I think, a hugely challenging problem. problem. I mean, you've got many really thorny issues at the heart of inventory optimization and supply chain problem. A number that you've talked about in previous talks I've seen you give along the lines of unpredictability of supply chain. So can you give us a sense? I think it's probably one of the biggies. So maybe we just dwell a little bit on that, on how painful and how difficult unpredictability can be for any tool that's going to purport to successfully manage an inventory in a complex supply chain.

7:27So that's exactly the purpose of inventory, buffering around uncertainty and buffering against uncertainty in demand and buffering against uncertainty in supply. Uncertainty in supply might be variability of lead times or alternative delivery channels or sometimes even serious events like the difficulties that we had a few years ago with the Swiss channel. And this is the reason why, in order to maintain service levels, we need to keep a certain amount of inventory in the network. So this is the purpose of it. And the fact that advanced modeling is able to reduce it is actually very positive because it demonstrates direct business benefits, the direct implication in terms of increasing or maintaining service levels while releasing working capital.

8:29But yeah, this is not doing that or doing that not at the competitive level translates in holding big inventory, being less competitive, holding working capital in our warehouses instead of investing it and so on and so forth. So, yeah, not being able to cope with those stochastic phenomena is quite impactful for businesses.

9:01Dr Lucy Morgan:Okay, so let's get into it then. So can you walk us through the technical core of that model and capturing that noise that you specifically talked to? Because I think that's going to really bring this to life. And what model did you use? How did you capture the noise? What were the inputs and outputs? And yeah, it'd be really nice to hear a little bit about the detail. So the A-Inventory Planner is an ensemble of several models. We can think of them as algorithmic components, if we wish. So first of all, we have a demand forecasting module. The demand forecasting module, when we started working on this project, was quite advanced, quite state-of-the-art in terms of availability of algorithms.

9:48Actually, we used global forecasting algorithms able to cross-learn between different time series in order to improve the performance. But the most relevant part of that forecasting was not producing point forecasts, but producing distribution of forecasts. Distribution forecasts built by different techniques, for example, out-of-sample bootstrapping to understand what is that distribution of our errors. because, as I mentioned, that was the critical point, being able to understand the distribution affairs and being able to buffer against the likelihood of being wrong, not against the likelihood of being right on a quite variable demand scenario.

10:33So that was the first component. Second component, as I mentioned, our task was to leverage the digital visibility that we already have in the network. So we created a delivery forecasting model, which was based on survival analysis. And we had different versions of it, like very simple, like a fit of variable distribution from a set of delivery times. or a fit of variable distribution where the regression model on the two parameters of the variable distribution can be even a deep newer network that captures all sorts of things like the season or certain specific atmospheric problems on our delivery routes and things like that.

11:14So different levels of complexity. So demand forecasting model, supply forecasting model. And that we had a, and we worked a lot on that quite interesting scenario simulation engine for our supply reality, right? We had a lot of historical data, which was part of the input, of how basically our good travel around the world, what are some specific milestones that are hit at certain point of the process. We don't have real-time visibility or we have partial real-time visibility and what happens with our purchase orders, but we have good visibility in the systems on when purchase orders reach specific milestones.

12:03So firstly, we use those milestones as intermediate steps or intermediate information in terms of upstream visibility used by the delivery forecasting model. And in fact, the delivery forecasting model was refining its predictions when the purchase order was going to be available in the destination, basically destination location. and was also gradually narrowing down the confidence range around that. And then in the scenario simulation engine, we are able to execute that hypothetical trip multiple times, multiple times compatible with all sorts of very complicated operational constraints. For example, we are able to place an ordinary order, what we call the ordinary order, within certain conditions of minimal order quantity, incremental order quantities, ordering frequency, make frequency.

13:05But sometimes we are also allowed to place what is called expedited order. But there are very serious rules around placing expedited order. The goods that you're ordering must be already in the production schedule. Otherwise, you cannot produce them. So our simulation engine, we created quite a sophisticated simulation engine that was able to simulate delivery and simulate that distribution and supply process in stochastic scenarios, in scenarios affected by stochasticity, but also compatible with that crazy amount of real world constraints. And as a result of that, we created a constraint optimization problem, which was basically, give us a business assignment.

14:01And the business assignment was, please, your policy, inventory management and ordering policy should be able to guarantee with a very, very high likelihood X service level. because this is basically the commercial constraint. We need not only commercial but also regulation constraint. We need to be able to guarantee certain service levels of our consumer health goods in all markets. So this is the business constraint. And then compatibly with that, we were selecting basically simplifying down to two parameters, the probabilistic trigger for a normal order and the probabilistic trigger for an expedited order when there was a high likelihood of having a stock out.

14:49And based on that, going through multiple simulations, we were able to draw the ISO service level curves and actually moving along those ISO service level curves, we were able to recommend the optimal policy, which was the optimal ordering policy for normal orders, expedited orders expected to guarantee a requested service level with minimum inventory.

15:21Dr Lucy Morgan:Wow, that's a heck of a journey. So you're starting with the raw data, you're seeding, you're parameterizing these fairly straightforward and very commonly used distributions, the ViBall distribution you mentioned that's often used in supply chain modeling with the data that you have surrounding your supply chain and you're using potentially AI models, machine learning models, neural network models to generate those parameters? Was I right with that? Yeah, yeah, to recommend optimal configurations of those parameters, optimal configurations for those parameters. And we actually executed multiple multiple shadowing exercises in which this AM entry planner was shadowing that the true process and was actually able to demonstrate the ability of reducing the total entry position by double digits 12, 13, 16, 15 percent while guaranteeing same or even better service levels And the funny thing is that we were calling it the M-ventry planner, but in this case, actually, it's quite pertinent because we observed that this ensemble of machine learning and mathematical components sometimes was finding solutions which were sometimes similar to what the real planner with the available tools was doing, but sometimes it was quite different.

16:54Sometimes it was reducing the frequency, sometimes it was increasing the frequency, sometimes it was timing the orders in a specific way. But it was like a non-trivial strategy, non-trivial order placing policy that this AI solution, that ensemble of these components was actually proposing.

17:10Dr Lucy Morgan:I guess, finally, I'd love to get on to one of the other projects you've been working on. But what does it mean then to scale that kind of planning tool in a global, multinational business like Helion? How did you both get the buy-in, but also then get the operational touch points into the business so that it could advise at every point of the supply chain process, potentially, for Helium? The full, very ambitious and fully automated version of the tool has not yet been fully adopted by the business, despite at certain point we even got the President's Medal. We were finally finished for the President's Medal of the Operational Research Society.

17:54However, the learnings or parts of our algorithms, such as, for example, the inventory corridors, the statistical probabilistic calculations of inventory corridors and execution, the simulation engine, the way in which we calculate planning parameters. So all that knowledge is now being implemented in a global enterprise program, which is our program on how we actually do integrated planning across the network. So all integrated business planning solutions and modern platforms, they actually rely on quite serious, quite powerful algorithmic components such as optimizers, simulation engines, a cascade of demand forecasting, a cascade of demand forecasting algorithms.

18:49So a lot of learnings from that project were actually directly implemented in the demand forecasting bit, in the scenario simulation bit, in how we compute the variability of inventory and are prepared for different scenarios. and how that process, that stochastically modeled process can then feed back in the actual planning process, not just in terms of parameters, but also in, for example, heuristics fine-tuning when you actually execute those planning algorithms.

19:26Dr Lucy Morgan:I wanted to move on to another project you have worked on with your team in the last few years, And this is equally amazing project involving the sort of manufacturing of your of your product. So I want to talk a little bit about the Golden Batch project. Can you give us a short sort of a pitch on what the Golden Batch project is and why it matters in pharmaceutical and consumer health manufacturing specifically? The golden batch approach in manufacturing and in pharmaceutical manufacturing is typically a way of formulating what is the optimal manufacturing pattern for a lot or a batch of products.

20:12And can be a combination of, for example, process specifications, regulation targets. The parameters of the batch should be within these limits. And also there is huge space for data-driven study. Sometimes when the process, the manufacturing processes are long, actually your control variables, your levers, your parameters that you can use to influence the process in real time or whatever time horizon you like, actually give you indirect information of what is going to be the final configuration or the final characteristics of the product. And this is quite the case with, for example, glycerol-based toothpaste.

20:59Glycerol-based toothpaste, which we manufacture, is an extremely complex non-Newtonian fluid that presents all sorts of challenges in the mixing stage. And the manufacturing process of glycerol-based toothpaste is very complicated. There are many different phases, like around 30 different phases, in which different things are happening, like ingredients are added, mixing happens at different speeds and temperatures and pressures, and then you have to cool down the product and so on and so forth. And the control variables that we have are the mixing parameters, actually the sensor readings that we get directly from the mixer.

21:42So we worked and found a way to actually influence and control that process through the mixing parameters to the sensor readings that we're getting directly from the mixers. And quite relevantly, the data-driven component of that golden batch construction, in our case, was quite heavy. Because of the complexity of the fluid, it was very difficult from a purely engineering perspective to design an entirely physical control system which is based on basic, on fundamental principles. So we had to very heavily rely on the data-driven part of that, analyzing thousands of batches, finding that optimal golden execution which realizes the best quality parameters within the shortest execution time.

22:40Dr Lucy Morgan:That's amazing. The challenge here is phenomenal. And so you've got 30 stages of potentially mixing, cooling, pressurization. How on earth did you manage to establish what the edges of your corridor of appropriate interaction was? the upper and lower limits of your levers, if you like, the rail builds. How did you manage to establish what that was at each stage and unpack it so that you could do that? Yeah, that was exactly the difficulty. Designing golden corridor envelopes, for example, if you have multiple realizations of the process, is relatively standard practice in engineering control charts, For example, you can build a statistical replica and understand when the process is drifting out of control and so on and so forth.

23:40But this is quite typical for stationary processes. In our case, we had a highly non-stationary process with multiples and they were in different duration. And it was difficult to recognize the exact time transitions directly from the data. So it posed all sorts of challenges. And we had to work really hard, for example, on the recognition of the step, the phase step, the phase change directly from sensor readings, which typically are very noisy. But yeah, eventually we got to this smoothened corridor, which was taking into account optimal transitions at optimal times. and eventually we actually got scientifically non-trivial results.

24:30We demonstrated that we're able to keep the viscosity of that fluid under control, standardize the viscosity of that fluid through that highly non-trivial and non-stationary golden corridor, which was entirely derived from data and from past and from simulated batches, from past and from simulated batches.

24:53Dr Lucy Morgan:Can you give us a little bit of insight then as to what the architecture for that actually looks like, what the sort of key engineering components were to bring that into reality? Yes. So this is a question which is interesting, not just from a scientific and from an engineering perspective, but also in a way from a business perspective. right because we started and we completely internally developed this project so we have to first of all demonstrate that it works having a working prototype demonstrating that whatever process we develop they're executable in real time or in near real time and at that stage of the project you rarely can say just buy me a platform to do that right so we have to end of use the tools and the environment and the components of architecture that we already have.

25:52And then probably at the next stage of the project where the prototype proves that it works, the algorithms prove that they're efficient, and actually there is some level of demonstrated business benefit at that point, you can think of more advanced solutions. And we used our available data science toolkit. So we had an execution environment where we actually trained and executed the models. We were able to interface that environment with what is called a data historian. Equipment, an asset platform, which we have in different sites around the world, that collects and stores and is able to actually interface with different systems, data streams that come from all possible engineering assets.

26:43So this is kind of probably slight alternative to having IoT devices installed on each single device, right? But our engineering assets produced a lot of data streams from all their sensors. And all these sensor readings were directed and managed through a piece of equipment, which is popularly known as data historian. So that was the first bit, how we get the data in the earlier time through the data historian into the execution environment and then provide the necessary visualization. And the visualization was done through a popular dashboarding tool that you can easily imagine. And yes, so this architecture worked.

27:25It was possible. It was quite sufficient from the viewpoint of proving the concept, but it's not a stable solution. It's not something that you can stably rely on. So once the business benefits can be demonstrated, probably thinking of deploying some of the compute power on some sort of edge technology instead of centralizing everything in one environment, having much better visualization tools that are more suited for the real-time visualization and refresh rather than having heavy dashboarding capabilities to do that. Yeah, so there's a lot of learning. One of the difficult points was actually, and it took some time, getting an architecture design approved by security because obviously there are a lot of firewalls and then security measures put in place on the manufacturing side so that external environment cannot be interfaced easily with process critical systems.

28:36So we had to do a lot of work to kind of, even with our prototype architecture, we had to face some of these serious technical aspects.

28:47Dr Lucy Morgan:And what I loved about the visualization I saw on this was that you called it a sort of manufacturing control corridor for the process. and then you were able to visualize it. Did I gather using some sort of principal components analysis? Visualize it actually as a corridor. It really does look like a corridor of sort of within this roughly cylindrical sort of hyper tube, you have to say. Yes, it was, in fact, the visualization was based on non-principal components analysis, then obviously evolving over time. But the visualization is kind of simple and nice, whereas the fact that we had to translate a highly non-stationary process to that seemingly stationary and nearly cylindrical hyper tube was an interesting part of the mathematical challenge.

29:53Basically, making a Schuhart control chart, which is meant to be applicable for stochastic but stationary processes, to make an adaptation, some sort of geometrical adaptation of it, to have an envelope defined by best realizations of the physical process. But yeah,