Machine learning has taken diagnostic, prescriptive and predictive analytics to the next level, sorting hundreds of data inputs to detect trends and problems preventing optimal yields.
Advancements in AI tools are enabling ethanol producers to optimize production, achieving insights more quickly and quantitatively, and with more precision. It’s being done with advanced data analytics, machine learning tools, holistic plant models and fermentation models.
IFF is at the forefront of the movement, according to Taylor Pellerin, technical service account manager, with its XCELIS® AI tool. “Helping out our customers the best way that we can has been the exciting part of being at the forefront of it,” he says.
While the ethanol industry has largely adopted AI-type sequencing technologies, other AI applications are taking longer to catch on, says Carson Merkwan, business development manager with Direct Companies. “Ethanol has had a culture of not wanting to be the first one, but being willing to be the third one,” he says. “Being the first one in is always more expensive, more risky.”
Still, XCELIS AI has been popular with customers and IFF is running models for a few producers curious about the functionality and benefits.
“There’s a lot of potential; there’s a lot of possibilities,” Pellerin says. “It’s an exciting field. I’m excited to be a part of it, trying to bring it to the ethanol industry.”
Models and Management
XCELIS AI is a customized machine learning tool that takes hundreds of inputs on a specific plant and spits out the outputs requested of it, perhaps ethanol or distillers corn oil yield probabilities, or overall plant efficiency.
“The goal of these models is to better understand how the process works and how to optimize the process,” Pellerin says. “For years now, the industry has been inching its way closer to that ethanol theoretical yield maximum, and so we’re hoping with the use of AI, trying to better understand those tiny little changes, that we can make that big difference to get us closer to our theoretical maximum.”
XCELIS AI has been used primarily in diagnostic applications but is expanding into predictive and prescriptive analytics.
All 200 or so features in the model generate a probabilistic outcome, Pellerin explains, so a producer can learn what changes would result in a better probability of higher yields: predictive analytics. “We’ll look at 90, 80, 70% chances of above-average ethanol—are any of these easy changes we can do right away?”
Prescriptive analytics allow a producer to learn what changes can be made to avoid past production issues. And diagnostics pinpoint the problem.
A production problem can occur with large changes, or small ones, Pellerin emphasizes, but always more than one. “It’s never only one thing. It’s going to be that one thing plus a few hundred others. It’s understanding the magnitude of each of those features and how each of those features change that yield for any given batch, for any given day.
“We’ve looked at things as large as the ambient temperature outside and how to optimize fermenters in the summer versus winter, to something as small as agitator amps on the mixing blade in the fermenter tank and seeing that even make a difference. It’s things like that where you wouldn't have normally put the time towards looking at it.”
XCELIS AI has modeled scenarios for plants looking to trial a new yeast product, exploring all the different levers that could impact plant performance. “The model is able to look at all those levers individually, as well as conditionally, grouping them together and seeing which levers we can pull that will help just get a little more yield out of that yeast that wasn’t there based on their current operating conditions.”
Going deeper, the secondary interaction of all those features could be 40,000 inputs and, further, up into the billions, Pellerin explains. “Something as simple as the liquefaction process, you can have as few as eight or nine features, square that and that's all the secondary interactions, cube it and you’re looking at even more interactions. It can be a lot.”
The model analyzes all the interactions and points producers in the right direction, he adds. “The nice thing about that is you can pick those low-hanging fruit and get those yield increases that are easy right off the bat. For example, by utilizing feature interactions, our models have shown success in fine-tuning liquefaction processes, predicting dextrose equivalent, down to fractions of a point to maintain consistent performance when upset conditions occur.
“As you go on, it’s an iterative process, you can see smaller and smaller changes and try to make changes based on those.”
Inputs and Training
Absolutely vital to machine learning is training the model, using inputs to help it learn how the plant operates, Pellerin says. The model will quantify and distill the inputs, pinpointing the most important inputs to the scenario being assessed. This type of machine learning is a subset of AI, like the large language models (LLMs) that are gaining popularity.
“But we’re not looking to make chatbots,” Pellerin explains. “We’re looking to make a black box where you input data … and the output is whatever you want the output to be. You’re looking at ethanol yield or corn oil yield—it doesn’t matter. Where IFF excels is evaluating multiple performance metrics to provide the most accurate model, then look inside of that black box to understand why the model provided that result.”
Predicted yield is compared to actual yield to verify accuracy of the output. “If it’s good, then we can move forward at interpreting that prediction and try to get insights out of it. If it’s not good, we adjust the methods and the weights, the way the inputs are being calculated. That’s training and validating the model.”
This type of specific AI learning customized to each ethanol plant is the next phase of AI’s evolution in the industry, Merkwan says. “You have a complex loop with a bunch of different inputs and outputs and you train your AI to reach a certain goal and you reward it when it reaches its goals.
“When you’re done training it, it’s ready and until your process changes, you don’t have to train it again. So that works out well.”
The data must be organized well, combed through and properly collected, he adds. “If AI is trained on bad data, it won’t turn out very well.”
Impact on People
The models, while adept at identifying trends and potential issues, do not suggest actions, Pellerin says.
“A human can look at an HPLC readout and tell you what kind of sample it was. The model doesn’t know what those numbers are—it just knows they’re numbers. It’s looking for trends, patterns and correlations and making those predictions.
“And for that reason, with machine learning models, we're not trying to replace subject matter experts. The goal is to give that subject matter expert the best tools available to make those decisions and be able to quantify those decisions.”
Merkwan, however, explains there are workplaces with no shift workers, what he calls “lights out.” Other industries are hoping to get there, even with the use of specialized robots, he says. “We have worked in other value-added ag industries, and they would like to get to as close to a lights-out facility as they can with zero shift employees onsite.”
But it’s not an achievable goal for ethanol plants, as they are now, he says.
“The way I see AI developing is it will be an informational-based thing. Robots are still expensive, even if you didn’t put AI on it, and you were able to teach the perfect robot, it would still be very expensive.” No industry is at that point yet, he says.
“You could get AI to tell someone what to do, and send them AI-generated videos, diagrams, SOPs [standard operating procedures],” Merkwan adds. “It could tell someone what to do on control systems, analyze their data and tell them what they’re doing wrong. That could be doable in the near future.
“But any time you have to turn a wrench on a bolt, you still need somebody, though that person might not be having to work as hard or to think as quickly because the AI might be doing a lot of the management part for them.”
Merkwan says he does not anticipate massive layoffs as AI advances in the ethanol industry. “Generally, when people are putting in lots of automation, it’s because they couldn't find someone to fill the position, not because they’re trying to replace people.”
The potential of AI has been better received in rural areas, where plants might struggle to hire a full staff. “When they are excited about it, it’s where the workforce in the area is not as robust,” Merkwan says. “They might be looking for five employees and they can’t get that. We’ll see that in very rural areas like North Dakota. People in those situations are very aggressive about automation and the opportunities for AI and robots.”
Erica Montefusco, senior vice president of risk and compliance for PROtect, touts AI’s benefits in drafting job descriptions when hiring. She explains the workforce has evolved past the initial core group of experts. “I think a lot of things kind of got stale in the ethanol industry for a while. I think you had a really good core group of people working in the ethanol plants following the build out of those facilities.”
When those subject matter experts leave their jobs at plants in rural areas, the candidate pool might not possess the breadth of experience in the technical nuances of those plants.
“When you have admin people hiring for those roles, it’s hard for those second-generation plant managers, operations managers, maintenance managers, to put into words clearly what skills they want people to bring to the table that makes them a good employee.”
LLM tools like ChatGPT can help clean up a stream of consciousness into a professional, clear job description, she says. And just like training a machine learning model in a plant, training an AI writing tool is crucial. The more information given, the better the outputs.
“It’s as good as the info that you put in, and I think that's where people often struggle with AI,” Montefusco says. “It’s truly garbage in, garbage out or quality in, quality out.”
Future of AI
As AI evolves further, it will become less expensive, more efficient and more integrated into the whole plant, Merkwan predicts. “People are figuring out better ways to prune the data back, use easier hardware, use less energy.”
In the near future, perhaps models will not just spit out numbers, but learn interpretation of those numbers, as well, he adds. “These will be custom apps where you dump their data into algorithms. There’s some time and training involved with that AI, but once that happens, they will be hitting the ground running and this AI can start taking control of these control systems.”
Pellerin says, “In the future, models will become faster, getting more real-time insight, they’ll become smarter at interpretation. But the one thing we also have to consider is, as these models get smarter and faster, they require more power to operate and more energy to train. So, we need to be conscious of our environmental impact as we expand these models.”
For now, Pellerin says IFF is focusing on expanding implementation in the industry through robust communication with customers to increase overall understanding of how the XCELIS AI models work and how to interpret them.
“We have to provide the correct explanation of this,” he says. “You can't just go up to someone with a chart and say, ‘here, trust me.’ We’re talking about real-world problems, real-world scenarios.”
This work dedicated to full comprehension of machine learning will help usher in new eras of AI.
“Understanding how AI models work and how the models can be interpreted will set people up for success later when we get to the point where AI is more integrated into the plants than it is now,” Pellerin says.
Author: Lisa Gibson
lisa.gibson@sageandstonestrategies.com