The short version. We partnered with Capital Gain Consultants, a German process-engineering firm, to speed up a slow, manual job: designing chemical separation processes for multicomponent mixtures under shifting constraints. Dot Square Lab built a machine-learning assistant that enumerates the feasible distillation sequences from combinatorics and engineering heuristics, then applies Bayesian multi-objective optimisation to find the configurations that best balance separation quality against energy, feedstock and material use, scored against DWSIM, the flowsheet simulator used as the benchmark. It outputs ready-to-open flowsheets an engineer can load straight back into DWSIM and take further, turning a task that took hours or days into one that takes minutes. The system now runs in production as the Flowsheet Copilot in Simulate 365, the commercial platform built on DWSIM.
At a glance
- Sector: process engineering (chemical, pharmaceutical, life sciences)
- Problem: manual, labour-intensive design of multicomponent separation processes, with inconsistent outcomes
- Approach: combinatorial generation of distillation sequences, then Bayesian multi-objective optimisation, scored against DWSIM
- Tech: Python optimisation pipeline (Gaussian-process Bayesian optimisation, building on HEBO), API integration with the DWSIM flowsheet simulator, flowsheet output loadable back into DWSIM
- Outcome: an engineer's assistant that returns optimal, ready-to-open designs; a task that took hours or days now takes minutes
- In production as: the Flowsheet Copilot in Simulate 365, the commercial DWSIM-based process-simulation platform
The challenge
Designing a chemical process to separate a multicomponent mixture is a combinatorial problem with expensive trade-offs. Engineers manually assemble unit arrangements (distillation columns and the steps around them) to hit separation targets while holding down energy, feedstock and material consumption. The constraints change from job to job, so the work is redone by hand each time. It is slow, it depends on scarce expert time, and it produces inconsistent material outcomes.
We partnered with Capital Gain Consultants, a German firm specialising in process engineering for the chemical, pharmaceutical and life-science industries, to automate the search for an optimal design, using DWSIM, a standard open-source flowsheet simulator, as the benchmark for what counts as a good design.
The approach
DSL built a machine-learning system that connects to DWSIM through an API and searches the design space in two stages.
- Generate the candidates. The system combines combinatorics with engineering heuristics (the "expert knowledge" a senior process engineer brings) to enumerate the feasible distillation sequences allowed under the user's constraints, instead of relying on someone to propose a handful by hand. Every viable arrangement is on the table, not just the obvious ones.
- Optimise across competing objectives. Each candidate is scored with Bayesian multi-objective optimisation. Multi-objective, because the goals genuinely conflict: higher separation quality often costs more energy or feedstock, so there is no single best design, only a Pareto front of trade-offs to choose from. Bayesian, because evaluating a design means running the candidate through DWSIM, which is expensive. A Bayesian optimiser builds a probabilistic model of how designs perform and spends each evaluation where it expects to learn the most, reaching strong configurations in far fewer simulator runs than a brute-force or grid search. Here that model is a Gaussian process fitted to how designs score, and an expected-improvement acquisition chooses the next design to run; the setup builds on HEBO, a sample-efficient Bayesian-optimisation method.
The system writes its results straight back out as flowsheets that DWSIM can load, so an engineer opens a finished candidate in the simulator, inspects it and takes the design further instead of decoding raw numbers. After a proof-of-concept validated the method against DWSIM, the optimiser was extended to cover a broader range of unit types and mixture compositions.
What we delivered
We delivered an engineer's assistant, not a replacement for one. It takes on the heavy part, enumerating and scoring candidate designs against DWSIM, and returns a short set of strong, Pareto-optimal configurations for the engineer to judge. Work that took hours or days of manual trial-and-error now takes minutes.
Just as important, the output is not a black-box number. The system writes out flowsheets that DWSIM can load directly, so an engineer opens a result, inspects it, adjusts it and takes the design further in the tools they already use. The engineer stays in control and sets the constraints; the assistant does the search. That is what makes it usable in a regulated, safety-critical engineering setting.
The system now runs in production. It ships as the Flowsheet Copilot, the AI assistant inside Simulate 365, the commercial process-simulation platform built on DWSIM, where process engineers use it to generate and optimise flowsheets.
Where this approach applies
The pattern generalises well beyond distillation. Any design or planning task that is combinatorial, judged on several competing objectives, and expensive to evaluate is a fit for the same approach of combinatorial search paired with Bayesian optimisation: heat-exchanger network design, reaction-pathway selection, formulation and recipe design, equipment sizing, and broader engineering-design optimisation.
The two ingredients that make it work are a trustworthy evaluator (here, DWSIM) and an optimiser sample-efficient enough to use that evaluator sparingly. If you have an in-house simulator or model and a design space too large to search by hand, this is a well-proven way to put it to work.
Frequently asked questions
Can AI design a chemical plant? Not on its own, and it should not try to. Used well, AI is an engineer's assistant: it takes on the heavy search a person cannot do by hand. In this project a machine-learning system enumerated the feasible distillation sequences and applied Bayesian multi-objective optimisation to find the designs that best balanced separation quality against energy and material use, then handed the engineer ready-to-open flowsheets rather than a black-box answer. The engineer sets the constraints and judges the trade-offs; the assistant turns work that took hours or days into minutes.
What is multi-objective optimisation in process design? It is optimisation when you care about several goals that pull against each other, such as separation purity, energy use and feedstock cost. Instead of one best answer it produces a set of Pareto-optimal trade-offs, designs where you cannot improve one objective without giving up another, so the engineer can pick the balance that fits the job.
Why use Bayesian optimisation instead of a grid search or a genetic algorithm? Because each design has to be evaluated by an expensive simulation. Bayesian optimisation builds a probabilistic model of how designs perform and targets its next evaluation where it expects to learn the most, reaching strong solutions in far fewer simulator runs than grid search or many evolutionary methods. When every run is costly, spending each one well is what makes the search practical.
Can this work with our existing simulation software? In this case the system integrated with DWSIM, an open-source flowsheet simulator, through an API and used it as the validation benchmark. The approach is designed to wrap around an existing, trusted evaluator rather than replace it, so your established simulation tools stay the source of truth.
Have a design problem buried in an expensive simulator?
If you have a large design space and a trusted way to evaluate it, the slow part is the search, and that is exactly what this approach automates. Strategy to production, one team.
