Our Client:

A German-based company specialising in process engineering for the chemical, pharmaceutical, and life science industries.

Situation:

Designing chemical engineering processes that can effectively separate multicomponent mixtures (e.g., natural gas, biofuel or crude oil) under varying physical constraints is a time-consuming and expensive process.

For our client, designing chemical processes usually requires human experts to develop complex systems,composed of different units that can be arranged in multiple ways to achieve optimal component separation,while minimising consumption of resources such as energy, feedstock, and construction materials. This approach is often slow and inefficient, with varying levels of material efficiency, cost, and chemical stability. In addition, there is the challenge of accounting for material, energy and feed costs.

The goal was to build an optimisation system to automate the design of optimal processes to minimise the resource requirements (energy, feedstock and construction materials) and augment the workflow for chemical engineers by suggesting potential chemical processing layouts.

Using the simulation software our client developed for running high fidelity process engineering simulations,DSL could benchmark our system and compare results to industry standard solutions that were the result of knowledge and expertise gained over decades.

Task:

DSL designed a machine learning (ML) system that integrated with our clients simulation software via an API. This system enabled the optimization, evaluation and iteration of designs for chemical engineering plants until the required performance requirements were achieved.

To generate optimal designs for chemical plants, our model used combinatorics and expert knowledge to generate all possible distillation sequences under user-specified constraints. Once a set of potential designs were generated, each was optimised using state-of-the-art Bayesian multi-objective optimisation methods to find the best configuration of the plant.

Effectively, our system explored potential sequences and adjusted variables to understand the impact, working towards an optimal design process. To support easy and stable integration, we developed a Python wrapper for the clients API which allowed human-understandable description of chemical plants, with underlying process calculations run using our client's software.

Once the PoC validated technical feasibility, the optimizer was extended to cover a wider range of units and processes involved for a broad range of mixtures, including natural gas and biofuel. The integration of our system with our client's software enabled rapid iterative improvements until the desired performance requirements were achieved, which was captured by an objective function and configurable within defined constraints of a specific process.

Result:

Our ML system enabled our client to automate the design of optimal chemical engineering processes, significantly reducing the time and cost required to achieve the desired performance requirements.