AWS ML Summit 2021 | Cultivating a company-wide machine learning culture at 3M
![AWS ML Summit 2021 | Cultivating a company-wide machine learning culture at 3M AWS ML Summit 2021 | Cultivating a company-wide machine learning culture at 3M](/pic/aws_ml_summit_2021_cultivating_a_company-wide_machine_learning_culture_at_3m/bi11RlViWkxrMk0_.jpeg)
[music playing] Digital transformation AI is changing the world. I hear it all the time at 3M. I’M David Frazee and I bet you hear it a lot as well. What are you doing about it? What am I doing about it? We are leading 3M’s culture change, pushing, pulling, hacking, nudging, driving the new technologies and the boldest of these new technologies I believe is machine learning.
Why? I think 3M only wins with its customers with innovative, cool solutions and I believe we need modern AI capabilities to bring these cool solutions to our customers in the future, to remain competitive, to remain cool, to thrive or maybe even just survive. Your point of the journey may be different, but perhaps we can learn from each other here today. It’s both a technical journey and a human journey. So let me start and talk a little bit about 3M from the human journey side. We have something at 3M we call the McKnight principles.
William McKnight who was CEO for the company for 20 plus years in the mid 20th century had one phrase that we like to talk about a lot at 3M. “Hire good people and leave them alone”, the human journey part of this. How do we inspire people and provide the things that they need to be successful? It’s a culture of innovation and leadership and collaboration that’s led to many of the innovations of 3M. Many of you know the innovation around the post-it note and how cool and innovative that was in its day, but let me talk about one that maybe you don't know about. A gentleman named Dick Drew when he was visiting the automobile factories back in the early days of automotive manufacturing was observing how they created two tone cars and what they were doing is they were taking plaster and newspapers and tried to create that fine line across the cars to give you that beautiful finish of the two tones.
Dick was an inventor and a scientist at 3M and thought “Hey, I see something here that if I go back to the offices at 3M and leverage our capabilities with sandpaper, because what is sandpaper? Sandpaper is made in long sheets of paper, we put sticky stuff on one side then we put sand or abrasive in that sticky stuff and then we make it very effectively and efficiently to solve a problem with trying to make production better with respect to grinding and polishing. And if I don't apply sand to the sandpaper, then I have a sticky surface that has a fine edge and perhaps I can lay that sticky surface against this car and create a very fine edge and pull that sticky piece of material off and leverage innovative capabilities for our customers, like we do with sandpaper”. And so, the cool part of that is that of course, it took us a while to figure out how to make that such that we put enough sticky stuff on to stay, but not too much that it would pull the rest of the surface apart. But the very cool part of that was this.
So imagine how it would be to ship sheets of sticky paper like sandpaper is often sold. Not really possible or feasible, so the innovation in my mind that’s the most interesting is how do you create a delivery mechanism for this sticky stuff called tape? And the innovation is tape on a roll. So before this time, there was no such thing as tape on a roll, so now we all know there’s thousands of kinds of tapes on rolls and we take it for granted about the technology and capabilities that this provides, as simple as it may seem today. So Dick Drew, inspired by a customer, created tape on a roll.
So what does tape on a roll have to do with culture and innovation and so on and today’s 3M? We talk about, at 3M, our technology capabilities using the model of the periodic table. We all remember chemistry from high school and college around elements and the unique ways that we can take elements, put them together and come up with new capabilities compounds that can create capabilities that we did not have before. And that really is the heart of 3M today, is the ability to create new capabilities using these technologies to solve real customer problems and so if we think about these technologies at our fingertips, including data science and AI, how do we bring this cool innovative technology and the culture that’s required to deliver that into 3M? So I think one of the places that we found success is to start with one of the places that 3M is strong and has been strong for over 100 years, in the factory. So I've got a couple of stories to tell you about what we’ve done using some machine learning capabilities in the factory. And I'll start with a project we call CloudMaker. It’s applicable in the context of how we actually manufacture things like tape.
Tape and film and we have web lines that can very efficiently create capabilities of stickiness and optical clarity and variety of performance parameters that our customers need and you’ll see in the diagram on the right that we coat that, sometimes we cure that, we use ovens and we have unwind and winding steps and so on and of course, when it breaks that’s a big deal. We do not want our webs to go down and so are there ways we can apply machine learning technologies to create predictive models about how we can prevent the webs from breaking and we’ve implemented this at a number of factories using machine learning technologies that you’re all familiar with to create those predictions and actually taking the warnings that it’s providing for us, stopping the line and correcting that such that we can continue with good production. Just an example of how we’ve started, if you will, in the factory, one of the core strengths of 3M. Another example of how we’re innovating in the factory is with our polymer production facility which is in Europe and what we’re doing there is we’re taking monomers and combining them in different ways to create new polymers and there are a large variety and combination of these monomers to create new polymers and one of the goals is to optimize the use of these large reactors. Like any reactor, bad things can happen when these overheat.
And we schedule the reactors today using what worked in the past, being very careful not to overschedule, to optimize both safety and production. So the sequence of how we prioritize what monomers and what reactions that we put together to create this production is something that’s a need for the company, because the less product we have, obviously the less output that we can get. So if we can increase the output, we can increase our ability to grow the company. So what we did was we took the data that we got from these reactors and did analytics against this using basic S3 and EC2, a little bit of SageMaker to implement this RNN model that predicts for the next 12 hours how the reactor temperatures and performance would be and using this prediction then we can create what if scenarios that allows the operator to then prioritize and perhaps create new time for reactions that we didn't see possible before. And this is how they would do that, they would look at the screen and they would do what if scenarios around putting certain reactions and certain production in different reactors over different periods of time to keep below the temperature limits and create new material production that takes advantage of all the capabilities of the factory. As we move forward with this, we aim to expand this using Amazon SageMaker further and then the newer services like time series analytics and of course we’re using the frameworks of TensorFlow and PyTorch to make this much more of a platform capability across the manufacturing lines in addition to where it is here in the polymer production.
So with our factory success, what does that mean for moving it further upstream in the way that we are intending and hoping to grow the company? So let me talk a little bit about some of the R&D programs and projects that lead us to provide growth for 3M. I’ll start here, we first started working with AWS on our 3M health information systems business, health care IT is still a data rich environment, but I’m not sure we get all the benefits that we all want to have from the information available to us as patients and certainly our customers as providers and how that can be utilized in ways to make all of our health care better. So what we’ve done in collaboration with AWS is use machine learning technologies, natural language understanding to enhance the record management as well as the billing cycles for our customers in many, many of our domains within the healthcare space and really leveraging this mission critical now capability of AWS for customers.
And so we’ve demonstrated with this case how our entire operation, entire operating unit of 3M can benefit from using machine learning capabilities and cloud technologies. So how can we expand this innovative capability to other parts of the company? Several years ago we created a new capability for customers. I have right here the Filtrate Filter, this is my own furnace filter with the sensor right here I'll talk about. This furnace filter has a sensor on it which allows us to measure the life of the filter. It uses sophisticated algorithms about the data of your home combined with weather data, location information, all the things you’d like to aggregate to better understand how well your air quality in your home meets your standards and the performance of your filter and what we found is that customers who have this have better air quality and they buy more filters, which is a great thing for us. Combining core 3M technology in the manufacturing of these filters, which filtration is a key, innovative lead product line for 3M with IoT and machine learning capabilities provides new capabilities for our customers.
Let me talk of another example, in the space of oral care. So in oral care, circa 2005 you see that crude interface of how to place brackets on teeth for straightening and aligning teeth and comparing that to where we are 15 years later, aligners are a much more popular way to do teeth alignment and ortodontia and 3M has a product of creating these aligners which you use every two or three weeks a new aligner and the interesting thing about this case is that every customer has a unique product when they receive that from 3M. Every aligner is different, it’s customized to you and your progression of your teeth and so one of the dilemmas that we faced here is how do we deliver those aligners more efficiently and more accurately to solve the problems of aligning teeth? Even two years ago it took us two to three hours to set up and analyze and create the series of aligners that you need to move the teeth. We’ve implemented new technologies in machine learning, we had surprisingly good results, because we used unsupervised learning and within one second arrived at the optimal path of treatment every step of the way and so this fundamentally changes our ability to deliver aligners, another innovation capability for our customers to get aligners much more effectively and efficiently. Going forward we have some opportunities to take one of our core product lines in films and create new capabilities for customers using metamaterials. So there’s a physics component to how we manage light with films and so what we decided to do was use the physics of white as the discriminator network.
The input model are the modeling parameters by which we want to constrain the film to create this metamaterial capability. So what we’re using is adversarial network capabilities to zero in on the ideal metamaterial to create performance of films that are unique and in high demand in today’s new products. And the example here I think is a head’s up display in an automobile.
Today we’re limited by a very small field of view and very small ability to have head’s up displays in automobiles, how can we create new materials that allow us to expand that field of view much wider to direct the light in a very specific way to optimize both the visibility for the driver, as well as what they want to see in the head’s up display. So we’ve been using machine learning, highlighting GN of course as I talked about to create this new metamaterial that today does not exist and our customers would be delighted if we could provide this at scale. One of the ways that we think brings a lot of these things together, everything from oral care to better furnace filters to new materials is the democratization of these capabilities. How does one in the labs around the world know that these capabilities exist and do not have to start from scratch every time, do not have to dig through the legacy information capabilities of the company to find out if this work has already been done so something we call the model hub has been built which is a portal for all of our scientists to go, are there models and simulations, both CAD models and structural models, physics models and AI models that have been created that we can already use and build on and add to the portfolio of technologies and innovations that we already have? I attended a conference recently where Anders had this quote “Every business model is getting digitally hacked”. I interpret that to say and the pandemic and we all see this vividly is that being digitally capable is a requirement going forward, including technology and business and the business model getting hacked as consumers in business and in homes want to consume their solutions from companies like us differently than they did even two years ago. So the culture of our company to take advantage of technology and adapt to these demands is an important aspect of what we’re talking about, so let me explain a little bit.
So building a bridge to the future we’ve characterized a program around this, around something we call materials informatics, which is a research domain for companies like ours that are trying to advance our ability to research and discover new materials and how to manufacture them and get them to the customer. So this cartoon if you will, is our roadmap that we’re using to get from one side of materials informatics, which is materials development with I call it old technologies, to new material development capabilities, maybe highlighted by something like this, the flow in the old days, in many ways today at 3M, down this river with lots of IP questions and datas and islands everywhere and lots of approvals necessary because things are not automated, into a future that has virtual capabilities and collaboration that’s seamless across different domains, different process spaces, different material spaces, collaborations outside the company that are seamless, this virtual materials informatic environment that we want to build for our researchers that would eventually lead to new products for our customers. This is one way we look at this. Our R&D community has always had data, we’ve always done R&D design experiments and we’ve created and automated the production from small sample sizes to large, but can we bring our data together and visualize the data that we already have? How can we vastly improve our ability to create simulation and predictive capabilities throughout the lifecycle of the process? And then perhaps most importantly, let’s bring automation to this, such that researchers can run experiments and learn the space that has the most likelihood for success of a new capability and the success for the next set of experiments to understand, does this solve the customer’s problem? That automation is not just automating what’s done manually, but creating new automation that allows us to do things that we could never do before, a really important part of that.
You will recognize many of these icons here about how we are on a journey from where we are today with islands of data that are not well integrated to the future of that virtual environment of materials research that I just alluded to, serverless capabilities, how do we put our data in lakes that are easily accessible and governed well? How do we use smart search and advance our innovator’s awareness of things that have already been done and new technologies that they can employ to solve the problems? Let me get a little more specific. Reimagining research in many senses to us could be something like this. Imagine that a customer comes to the offices at 3M and they talk about a hard problem that they’re trying to solve with their new products and services that they want to provide and said “I really want to know 3M, if you can build a new adhesive or a new abrasive or a new film or a new filter, filtration capability that would allow me to do this, this and this” and today we would have our scientists get together with their scientists, we’d set up some follow up meetings then we’d do some evaluation and some analysis, do a pilot run and perhaps send them a sample in four weeks, six weeks, eight weeks. And that works and that works well because we’re known for being able to create these new products based on some pretty challenging requirements, but imagine maybe a way that we have the people in the room talk about the problems, we go to the portal to map out what solution space this customer wants to have this new technology and we start the simulations and the experiment immediately because of things like Model Hub and we start dialoguing with the customer while they’re still here and say “Well we can make these three or four things, get feedback on that and then actually build them, run them, make the prototypes, run the extruder, do the 3D printing of the materials” and before they leave town they have the samples. So we’ve taken an eight week process and maybe gotten it down to an eight hour process.
And what we’ve done and the way that we could do this is materials informatics, automation and now something that we like to call the scientist social network. How do our scientists know what’s available to them even in the what of the moment with the customer? Let me segue into something that we’re working on today, which we think tries to bring both the technical and the human side of doing innovative things together. So this is a map of a prototype of how one correlates the work product of an individual to clusters of work products and individuals together on a diagram. One of the dilemmas we all faced here is the fact that we could ask our researchers to go to a database and fill out the form of their research and their publications and their university experiences and their travel and their conferences. I think you and I all know that this just doesn't work.
People are not willing or have the time or take the energy to update their data on a regular basis, so we need to get this as part of the natural course of work that our scientists do. So there are so many data sources in my company, in your company that are natural parts of the work that we do. So for example, imagine “Show me people with more than 20 inventions in a certain category and their connections in the services” and then you shrill down to find out what is the invention and how are they connected. Here’s another version that we’ve created and it’s the “Show me the software papers published in the last 10 years” and you see in the orange bubbles initials of people’s names and how they’re clustered together, you see other relationships to different parts of the organization, different companies and we can time window this such that we can see how these relationships and clustering changes over time and if one can imagine going forward, all the technologies of that periodic table that we showed earlier an ability to drill down in real time as we talked about in that customer meeting. “Show me the filtration capabilities of this type of technology” and I can immediately find out who and where and what time frame and then I can start to immediately interact with that individual or that technology even while we’re talking to a customer. How cool would that be? And of course, there’s some information technology that underpins how this works.
We’re using Textract and Neptune and all sorts of new services from AWS to really make this capability come to light and come to life in ways that will bring people in, because the culture of 3M like you culture is probably already wanting to be collaborative, wanting to share, but how do we do that? How do we cross into this new realm where everyone has access to this in a way that brings excitement, passion, interest and belonging to what we’re trying to do and the hard problems in front of us? This is a slide from maybe three or four years ago. We had a big AI summit in Europe and we had all the right people together in one place. The problem is it’s a big summit, in person, in Europe, which these days is very hard to do, is not very inclusive, because others should be apart of the summit but are not able to do that, so how do we bring that technology and collaboration together with the culture and the staying relevant and thriving in today’s world? So one of the things that we found very successful from two summers ago now was a hackathon.
Started as a bit of an experiment to try to see how we could collaborate with each other using cloud, using some customer challenges in front of us. You see our CTO in the picture next to the deep racer track. This was a wild success and inspired us to continue to do more of these around the moniker of the curiosity code. And so we believe that this is one of the ways, one of the vehicles by which culture and collaboration and technology come together to provide the new. Last summer we did this hackathon and we thought “Oh, I’m not sure this is going to go, we can’t all be together while working from home” and we even surprised ourselves. It was bigger, we had 800 people from 19 countries hacking virtually together to again innovate and collaborate and the fundamental thing that linked us all together is cloud and the desire to be innovative for customers.
2020, 800 people, hundreds of projects company wide, this photo is from 2019 when we were all together, we hope to be together doing this more as well in the future. This is a great challenge to me and our leadership team. Why don’t we work like this all the time? We get together, we collaborate, we get inspired, we find new ways to connect, why don't we do this all the time? And the answer is well, we do.
We’re doing more hackathons, we just had a winter hackathon where a smaller number of people got together without all the pomp and circumstances of the big event and the big challenges of bringing hundreds and hundreds of people together and orchestrate and do all the mechanics of that. So let’s make this a natural part of what we do, a natural part of how we collaborate and bring that technology journey of how do I improve and learn and keep up with the human journey of how do I stay connected to my colleagues, both inside and outside the company? And I like the way that we sum up research reimagined here in a way that it’s a flywheel. So we have cloud which gives us scale, we have agile and scrum and rhythm of being fast and adaptive and with that we have people that are passionate about what they do and they bring that both technical capability, but also their desire to do something meaningful for our customers and for the company and create this culture that’s kind of like a hacking culture all the time. And we think that machine learning and the ability for us to collaborate with technologies that allow us to see what everyone’s working on in a constructive way, inspiring for everyone is a thing that can make the difference about how our culture can adapt to the new ways that we’re going to innovate for our customers. I thank you for your time. 3M science, applied to life, applied to culture, the journey’s technical, but the journey is human.
Thank you!
2021-06-19 18:06