.png)
APE-0 (Beta) : A Tool for Conceptual Process Design
What the future of chemical process design should look like.
Designing Chemical Processes is Hard
The development of new or revamped chemical manufacturing projects is tough — they often take 5-10 years, and hundreds of millions of dollars to complete [1]. Often, the introduction of new chemical processes, feedstocks, or technology jeopardizes the success of a project. Data shows that 35-65% of major chemical projects exceed their budgets by more than 25%, and have their schedules delayed by 50% [2].
In the best case, before building a big chemical project, developers rely on experiments or historical data to ensure success. This is the concept behind moving from bench to pilot to demonstration to commercial facilities; take your learnings from smaller scales to build effectively at the largest, profitable scale. Unfortunately, it’s impossible to run the exact experiment before construction begins. Even slow-and-steady scaleup approach is costly and slow, and often infeasible for smaller companies. This is where simulation is critical.
Simulation allows us to virtually test the process and its variations and to understand how the scaled-up process will behave. This allows us to understand techno-economics for cheap — simulating a plant is orders of magnitude cheaper than building it. But while cheap, simulations aren’t free. A single yearly license can run $10k-$100k/year; and it can take multiple engineers working for a year to reach sufficient fidelity... assuming the process is well enough understood to model readily!
In addition to these expensive licenses, teams still face:
- Steep learning curves for novices, who may avoid simulation altogether due to the challenges of mastering the software, and can make costly mistakes out of ignorance.
- Hurdles for experts, who often struggle to get systems to converge and to scale their experimentation because of software limitations.
GenAI for Chemical Process Design
At Alkali, we're building the first AI-Copilot for chemical process development. Modern Generative AI offers a unique opportunity to lower the floor for newcomers, and raise the ceiling on what's possible for experts in process modeling.
Large Language Models (LLMs), trained to simply predict the next word, are not great at math. You wouldn't want to directly ask an LLM to predict the final purity from a train of distillation columns — a result determined by a mesh of equation matrices.
Similarly, you wouldn't want to ask a seasoned engineer, away from their desk for lunch, the same question, and trust their prediction to high precision.
Like a human engineer, you'd want AI to use a tool, like a model, or a simulator, to find the answer. But, while it can take human engineers months, if not years to become proficient in software, an LLM can theoretically become proficient in software, which is just code, in minutes. AI is really good at understanding code.
Under the hood, APE-0 uses tried-and-true open-source packages to run simulations. You don't have to trust that the LLM predicted the result — you can check out the simulation file it produced, and run it yourself!
For Beginners
APE-0’s web interface lets you start simulating without wrestling with software quirks. Describe your idea (with words or an image), and APE-0 will build an initial model you can tweak, or use as-is for quick, ballpark results. Our AI is trained on a growing set of increasingly realistic examples, so you won't have to start from scratch.
For Experts
APE-0 accelerates setup and convergence troubleshooting, and it can explore parallel design variations overnight, freeing you to focus on high-level engineering decisions.
The Future
As LLMs march forward to provide value across professional disciplines, we imagine the future of engineering will look different. AI can help us explore ideas much faster, and inspired by a trove of past designs, in a more effective and data-driven way. We imagine this will help scientists the most — those who may not be experts in a particular software, but have a fundamental grasp of the underlying science — to pursue their ideas further into commercial viability.
GenAI for process design is extremely early. When you're building physical infrastructure, hallucinations could be costly. That's why (for now), AI should be used to speed up work with simulators, not replace them. AI promises to speed up development and produce better, safer processes, if we can nail the form factor. While AI can work tirelessly through the night to explore a massive design space, at the end of the day, designs should be approved by engineers. One day, skipping an AI audit of issued-for-designed plans, might feel as unwise as driving a car without lane-assist.
APE-0 is in early beta. There’s plenty to do before it can deliver fully realistic systems with accurate costing. This includes:
- More Unit Ops: Currently, we work with simple distillation columns, absorbers, and pumps. Contact us if you have a favorite unit op you'd like to see in APE-0.
- Accurate Cost Estimates: Currently we use Lang factors for equipment costs, and typical utility rates for operating costs — this produces very rough estimates — our cost models are always improving!
- Robust Convergence: Sure, setting up a simulation can be annoyingly time-consuming. But getting it to converge can be a nightmare! APE-0 can do early convergence troubleshooting, but we are just getting started!