Corteva, Inc. (CTVA) Earnings Call Transcript & Summary

Hello, everyone, and welcome to today's virtual session on accelerating value creation with Aspen Fidelis and artificial intelligence. My name is David Diehl, and I'm a Global Reliability Engineer and System Performance Modeling Subject Matter Expert at Corteva Agriscience, within our reliability and maintenance technology group. We'll start the presentation with a couple of slides about Corteva Agriscience, and our system performance modeling program. And then we'll take a look at the power of Aspen Fidelis when coupled with machine learning models and other business intelligence tools and how that can dramatically reduce the resources required to develop models and also the durations from project conception to business decision-making. Corteva Agriscience, officially launched in 2019, bringing together the rich heritages of 3 iconic U.S. businesses, 2 Fortune 100 companies, DuPont and Dow Chemical and the groundbreaking hybrid corn company, which was later renamed to Pioneer. These 3 companies brought to the table a combined 432 years of agricultural expertise and business experience. At Corteva Agriscience, our purpose is to enrich the lives of those who produce and those who consume ensuring progress for generations to come. Our global scale and market presence extends to roughly 140 countries around the world with a global workforce of approximately 21,000 employees. We have 150-plus R&D facilities, roughly 100 production and manufacturing facilities producing over 65 active ingredients, which impact 100-plus different crops around the world. We're headquartered in Wilmington, Delaware with global business centers in Indianapolis, Indiana and Johnston, Iowa, and various regional business centers around the world. In 2020, we had global net sales of $14.2 billion and recorded an operating EBITDA of $2.1 billion. Corteva's presence as a pure-play agricultural company allows us to provide the unique solutions for farmers around the world to maximize yield and improve profitability, which is paramount in ensuring a sustainable abundant food supply for a growing global population. At Corteva, the system performance modeling function within reliability and maintenance technology exists to serve the purpose of enhancing data-driven decision-making to ultimately enable business success. And really, there are 3 key pillars that align up to this purpose, being capital projects, mature growth molecules and the tools and processes, which support building and deploying Aspen Fidelis models. For capital projects, both greenfield and brownfield projects, the target is to deploy system performance modeling on all projects over $2 million and those being projects that are applicable to Aspen Fidelis and the types of opportunities that it can identify. The objective for capital projects is, in general, to identify opportunities to reduce CapEx by an industry average of roughly 5% or in addition to ensure the ability of the capital project team to meet the long-term business objectives of the project in terms of the future production capabilities. Most of the models targeted at capital projects consist of either high-level sizing or throughput models, which aid in early design decisions. They can also progress through to detailed design and spares, operational, maintenance or turnaround strategies. One of the single greatest barriers to implementing system performance modeling on a capital project or also a mature growth molecule or resource constraints. And generally, that is the internal staffing or the internal resource requirements to build and develop the models. Complex models for large projects might often include 500 to 1,000 or even more various different types of events, which define the reliability of the system, including things like equipment failures, supply chain-related events, scheduled maintenance, weather events and many others. And for each one of those events, the appropriate amount of data is required to be collected to characterize the failure rates and also the repair characteristics for each of those events. And doing that often takes a tremendous amount of time from a dedicated pool of resources. And that could be anywhere from several weeks to several months' worth of data collection. So this leads us into the tools and processes, which support delivering on those key modeling projects. And really, the primary objective of this third pillar is to remove the barriers to implementation for delivering modeling projects and also to create and deploy innovative solutions, which allow us to cater Aspen Fidelis to our specific corporate needs and the way that our processes tend to operate. So with that introduction, I'd like to move on to the primary topic for today's webinar, and that is accelerating value with Aspen Fidelis and artificial intelligence. And really, in addition, it's the end-to-end workflow using Aspen Fidelis in combination with front-end and back-end supporting applications to improve the way that not only we can develop models with better efficiency, but also we can enhance the reporting capabilities of the results within Aspen Fidelis to drive better business decision-making. And the first step in this process is data extraction and preparation. This is generally a very manually time-intensive part of developing any model. And it begins with extracting information from SAP, typically work history, asset by asset for each item that's going to find its way into your Fidelis model. And it's going through all of that data and classifying the work history as either generally failures or not failures of some piece of equipment or at a more detailed granular level, what specific failure mode has occurred. And that information serves as an input to create the frequency distributions or duration distributions, again, that are inputs into your Fidelis model. The next stage in the end-to-end workflow, of course, is Aspen Fidelis, and that really is the heart of the system reliability analysis. One of the things that I enjoy most about the Fidelis software is the flexibility of the application and the adaptability to each companies or each facilities or projects, specific needs. And those are represented there with the 3 different icons. The BP stands for batch processing. The P stands for production and the DE stands for data exporter. We'll go into those 3 in a little bit more detail in a few slides. But for now, know that their visual studio add-ins that have been developed in-house to better suit the specific needs of our batch processes and our various applications at Corteva, the data exporter of which feeds into the last of the end-to-end workflow, which we won't talk about today, and that is exporting the native results within Aspen Fidelis and some additional business-specific metrics to be able to use them or to view them from some business intelligence tool, in this case, we use Power BI, but there are many others available on the market in order to do that. And so there, we have a high-level overview of the end-to-end workflow to enable Aspen Fidelis and it starts with data extraction and preparation following through with modeling and then analysis and business insights. And for the next few slides, we're going to focus specifically on the data extraction and preparation, which consists of the extraction of information from SAP, the work history analysis, the frequency and duration, distribution, creation and the subsequent import into Fidelis for the work history analysis component of data extraction and preparation, we're going to take a bit of a deeper dive into how deploying artificial intelligence or machine learning models can help to dramatically reduce the amount of time that it takes to complete that operation, and likewise, also greatly reduce the resource requirements that it takes to analyze and classify your work history for inputs into an Aspen Fidelis model. So if we recall back to the barriers to implementation for modeling in support of capital projects or mature growth molecules, is the data gathering for complex models often takes weeks or months' worth of work. And due to the nature of the data extraction and the analysis and the creation of the distributions, it's generally a very manually intensive task. And this is in conflict with the fact that project teams need to make decisions with agility and also that resources are typically very constrained both from the project perspective and also from the operations perspective. And with these things in mind, models risk not being able to deliver the right solutions on time or alternatively and unfortunately, the decision might be made to simply not perform modeling in support of the capital project or to close gaps for mature growth molecules. And we immediately recognized this as a barrier to implementation for system performance modeling in a workforce where we're constantly forced to try and do more with less. The thought of not performing modeling on capital projects or mature growth molecules or for spare parts warehouses or for sites or for other projects, was simply not acceptable. And there needed to be a solution to address the problem of limited resources in order to complete the time-intensive tasks in the data gathering and extraction phase for building the inputs into the model. And really, it only made sense that the right solution would be some sort of AI-driven machine learning model to perform the same operations that a human was doing in classifying the work history, but to do it autonomously and with some minimal human involvement, but in doing that, significantly reducing the amount of time that it takes to develop the inputs that go into a Fidelis model. And so in the next couple of slides, we'll go into why a machine learning model is necessary or right or proper solution for this problem and just exactly how it works. So to highlight some of the complexities in classifying work history in a manufacturing environment, And also to show the need or the applicability of a machine learning model to be able to manage this task, I'd like to use an example of driving the car in your daily life. And so imagine there are 2 different scenarios. And in each scenario, your tire pressure sensor might alarm for low pressure. However, in one scenario, maybe it's a really cold morning and due to it being a really cold morning, the pressure in your tire is lower. And before you start driving, it's low enough that it's below the lower limit threshold and your TPS sensor alarms. And imagine in the other scenario, you might have driven over a nail a day prior, and you've got a slow leak. And so of course, you're actually losing air pressure in your tire and your TPS sensor alarm's low. In both of those scenarios, you might decide to call and make an appointment with the shop in order to have the tire inspected and checked out and any corrective action to have taken place. But in one scenario, even though the appointment's made and the car is taken to the shop, it's really just a false alarm, it was a cold morning. And as you started driving the pressure in the tire rose, and now you are no longer below the lower limit threshold. Maybe the tire just simply needed some air put back in it. But in the other scenario, maybe you, in fact, had driven over a nail and the tire either needs patched or replaced. In manufacturing, these exact same scenarios exist, except with manufacturing-type assets. In both cases, you might see a work request being put in to repair some piece of equipment. But maybe in one instance, it's really just a false alarm. There were some environmental conditions or operating conditions, which prompted a work request to get put in, but under further inspection and advanced troubleshooting, maybe there wasn't actually an issue with that piece of equipment. But the work request or the work order still exists in the system. And imagine it being able to tell the difference between a failure of some asset or not a failure of some asset ultimately leads to being able to collect the right information to feed into the failure distributions, either as generic or bispecific failure mode that ultimately feed into your Aspen Fidelis model. Now if you're working with only a few select assets and you pull the last 10 years' worth of history, you might only have 10 or 20 or 30 different records to look through, and that's a very manageable operation to do. And it takes very little time to go through each of those and look at things like cost or the descriptions that were written by the technicians in the order or even the emergency or the priority of the order. But imagine for a Fidelis model, especially complex ones where you might have 500 to 1,000 or even more assets in a model, and each one of those assets might have 20 or 30 or 40 or more records over a certain period of time. Now you're looking at tens or hundreds of thousands of rows of data that need to be analyzed and classified in order to create your failure distributions. And that right there is one of the single largest barriers to not only creating models with speed and efficiency, but getting the proper resources allocated to dedicate the amount of time that it takes to go through the process of gathering the information to develop new models. So to illustrate this a little bit better, we're going to use a theoretical example of a work order that's generated for the replacement of a centrifugal pump seal. So here, we have 4 fields, which define the work order. We've got the title, Replace PE123 Pump Seal. We have a description, which is entered into the system by an operator or a technician. We have a date, January 1, 2021. We have a responsible person, Jane Doe, and potentially a user ID. We've got a phone number. We have some generic headers, task details, problem description and PE123 solvent return pump, seal not leaking. We've got a cost of USD 80 and an emergency priority. So this is the information that we know about this work order. And now it's up to us to determine whether or not this represents a failure of some piece of equipment, in this case, the pump seal failure. Or if this work order is more of a false alarm, it's not a failure of some piece of equipment. And this is where we deploy the machine learning model to help us understand how to classify this order. And so to do this, we have a training repository where we've got 10,000-plus previously classified work orders. One is just like this, where we went through and we've said, this represents a failure or this represents not a failure of some piece of equipment. And so the machine learning model uses that previously classified information to make determinations based on different fields in the order or different text descriptions to understand if the unclassified work order is in actuality, a failure or not a failure of that piece of equipment. So it will look at different things in the value fields, such as the emergency priority or you might also have an alliance to a turnaround priority or required shutdown priority. But in this case, an emergency priority, you've got a cost. So you had -- USD 80 was the total cost to perform this job and $80 is essentially nothing in a manufacturing environment. So that's a little fishy. But in general, between the priority and the cost, there really isn't enough information to tell. So the machine learning model at this point in time isn't necessarily confident in its prediction. But we combine that then with the natural language processing or the text prediction component of the model, and we look at things like the title and also the description. So we do this in a couple of different ways. We look at 1 gram and 2-gram predictions. And so 1 gram really is just one word at a time and using feature selection to determine the importance of each of those words and being able to predict a failure or not a failure. And in this case, the 1 gram might pick out seal, and it might pick out leaking. And based on that, it might say, well, this is obviously a seal failure. There's a leaking sealant needs to be replaced. When you look at the 2 gram or the 2-word pairs in the text prediction, you see it a little bit differently. So you see seal not and you see not leaking. And that very clearly is an indication that the seal wasn't leaking. There wasn't actually a problem with this particular pump. The work order was maybe a false alarm. Maybe there was some debris caught in the seal phase and that depresses flushed out and everything is back to normal working conditions. The point is that these 2, the value field prediction and the text prediction are able to work together to ultimately provide a fused prediction and the subsequent output as a failure or not a failure of some piece of equipment. And now imagine that for all those thousands of work orders that you need to classify, you have the order, the asset and the date at which that asset either failed or not failed and also what the failure mode is. And those provide or serve as critical inputs into the creation of frequency distributions and from a different aspect, the duration or the repair distributions as well and as inputs into Fidelis. And now we can see the true power of deploying a machine learning model to assist in the data gathering and preparation phase and in some cases where it might take upwards of several months by one individual or a few individuals to extract the information and develop the frequency distributions. We can now shave that amount of time down to just a matter of a few days or a week and consider the benefits to a capital project team or a facility and being able to have the recommendations from a Fidelis model 6 weeks before they would have originally had that information, had the work needed to have been done manually. And so this concludes the section of the presentation focused on machine learning capabilities and how we can use automation to dramatically reduce the amount of time that it takes to develop Fidelis models and deliver business solutions. And so in the last slide, I'd like to go back to a comment I made earlier about the customizability of Aspen Fidelis and the true value that, that represents and being able to create specific solutions to address certain business needs. And if you recall, all the way back at the beginning of the presentation, when I was providing an overview of Corteva Agriscience, much of our crop protection manufacturing assets are batch processes. And so one of the first things that we identified with Aspen Fidelis was the need to be able to represent our batch process operations in the software. The benefit of Fidelis and Visual Studio is that with enough imagination, you can do this. And so we created an accurate simulation of batch process operation within Fidelis, including cycle time variability and step-dependent failure impacts, including whether the batch is in process or forwarding or receiving a batch. And that's proven incredibly helpful to articulate the needs of our batch processes where otherwise we might have tried to simulate them as being a continuous asset. The next example is the production simulator, and this is an approximation of future unit production using a production liable and historically demonstrated run rates as a baseline. And this is incredibly helpful in simulating a unit's performance in a site or an envelope model without the need to create detailed unit-level reliability models. And the last slide I'll share with you today as hinted at before, is the data exporter. And the data exporter within the user code collects the native Fidelis results and other business-specific results during the simulation run and exports it to a database for insights in business intelligence tools. And this is incredibly helpful for folks that either don't have access to the Fidelis software or are not familiar with the Fidelis software to be able to view the recommendations from the model in a platform that they're familiar with, whether it's Power BI or Tableau or some other BI platform. And the reason I wanted to share that with you today is to give you a sense of what's possible with Aspen Fidelis and user coding. And also back to the end-to-end workflow with machine learning and artificial intelligence through Fidelis, through the business intelligence software. And holistically, when included as an end-to-end workflow, the value that, that's able to deliver to a company. I'd like to thank everyone for joining today's session on accelerating value creation with Aspen Fidelis and artificial intelligence. I hope you all found the session informative and enjoyable. Thank you, and enjoy your remaining sessions.