ADOPT - Advancing optimisation technologies through international collaboration
Lead Research Organisation:
Imperial College London
Department Name: Chemical Engineering
Abstract
The complex, interconnected and fast-changing nature of today's society presents a growing challenge for decision-makers. Increased competition in the process industries (oil and gas, chemicals, personal care products, food, pharmaceuticals and agrochemicals) means that agility must be built into process design and operation. Furthermore, the need to ensure reliability across the supply chain, minimise resource use and environmental impact, and maximise energy efficiency combine to make investment and operational decisions especially difficult. Such multifaceted decision-making has long been aided by detailed mathematical models of physical and engineered processes, which enable digital twins and constitute a cornerstone of smart manufacturing technologies and the future Industry 4.0. But the full benefits afforded by these models have so far been hampered by the lack of tools for exploiting them beyond "what if?" scenario analysis. In particular, the uptake of optimisation-based decision-making has been hindered by the large-scale, nonlinear and uncertain nature of these problems that often leads to suboptimal or even unphysical solutions.
In the ADOPT collaboration between the Sargent Centre for Process Systems Engineering (CPSE) and the JARA Center for Simulation and Data Science (JARA-CSD), we propose to address some of these shortcomings by developing improved methods for deterministic global optimisation, a class of optimisation methods that rely on complete search techniques and offer a rigorous conceptual framework to overcome the caveats of local optimisation. Our key research hypothesis is that the integration of deterministic global optimisation with surrogate (simplified) models and machine learning will enable transformational changes in our capability to tackle complex decision-making problems, leading to more tractable solutions with global optimality certificates and improved resilience to uncertainty. This nascent area brings about the following specific research challenges that we shall tackle within ADOPT:
- identifying best-in-class theoretical / algorithmic global optimisation frameworks and surrogate modelling paradigms to empower surrogate-based optimisation;
- handling uncertainty within the chain linking physical/simulated data to surrogate models and to optimisation results; and
- developing bespoke deterministic global optimisation approaches for more challenging classes of problems beyond mixed-integer nonlinear programming.
The ADOPT collaboration brings together two world-class teams of researchers in the field of deterministic global optimisation as well as team members who are specialists in handling uncertainty, in solving large-scale combinatorial problems, and in applying optimisation to real-world engineering problems. Furthermore, our assembled team partners with prominent optimisation software and process modelling companies in order to increase the accessibility of the research outputs and facilitate their dissemination.
The ADOPT collaboration creates added-value through the combined strength of scientific expertise of the two centres, the breadth of the software infrastructure that can be brought together, the wealth of its human capital, the reach of its industrial relationships and the exceptional potential to establish a long-term partnership. It will lead to scientific advances that can be tested on practical problems quickly, ensuring maximum impact from the research.
In the ADOPT collaboration between the Sargent Centre for Process Systems Engineering (CPSE) and the JARA Center for Simulation and Data Science (JARA-CSD), we propose to address some of these shortcomings by developing improved methods for deterministic global optimisation, a class of optimisation methods that rely on complete search techniques and offer a rigorous conceptual framework to overcome the caveats of local optimisation. Our key research hypothesis is that the integration of deterministic global optimisation with surrogate (simplified) models and machine learning will enable transformational changes in our capability to tackle complex decision-making problems, leading to more tractable solutions with global optimality certificates and improved resilience to uncertainty. This nascent area brings about the following specific research challenges that we shall tackle within ADOPT:
- identifying best-in-class theoretical / algorithmic global optimisation frameworks and surrogate modelling paradigms to empower surrogate-based optimisation;
- handling uncertainty within the chain linking physical/simulated data to surrogate models and to optimisation results; and
- developing bespoke deterministic global optimisation approaches for more challenging classes of problems beyond mixed-integer nonlinear programming.
The ADOPT collaboration brings together two world-class teams of researchers in the field of deterministic global optimisation as well as team members who are specialists in handling uncertainty, in solving large-scale combinatorial problems, and in applying optimisation to real-world engineering problems. Furthermore, our assembled team partners with prominent optimisation software and process modelling companies in order to increase the accessibility of the research outputs and facilitate their dissemination.
The ADOPT collaboration creates added-value through the combined strength of scientific expertise of the two centres, the breadth of the software infrastructure that can be brought together, the wealth of its human capital, the reach of its industrial relationships and the exceptional potential to establish a long-term partnership. It will lead to scientific advances that can be tested on practical problems quickly, ensuring maximum impact from the research.
Organisations
- Imperial College London (Lead Research Organisation)
- University of Leuven (Collaboration)
- University of Wisconsin-Madison (Collaboration)
- University of Connecticut (Collaboration)
- IMT Institute for Advanced Studies Lucca (Collaboration)
- Eindhoven University of Technology (Collaboration)
- Octeract (Project Partner)
- RWTH Aachen University (Project Partner)
- Gurobi Optimization, LLC (Project Partner)
- Siemens Process Systems Engineering Ltd (Project Partner)
- MOSEK ApS (Project Partner)
Publications
Borovykh A
(2022)
Data-driven initialization of deep learning solvers for Hamilton-Jacobi-Bellman PDEs
in IFAC-PapersOnLine
Borovykh A
(2024)
Stochastic Mirror Descent for Convex Optimization with Consensus Constraints
in SIAM Journal on Applied Dynamical Systems
Bounitsis G
(2022)
Data-driven scenario generation for two-stage stochastic programming
in Chemical Engineering Research and Design
Campos J
(2022)
Partial Lasserre relaxation for sparse Max-Cut
in Optimization and Engineering
Chen Z
(2023)
On approximations of data-driven chance constrained programs over Wasserstein balls
in Operations Research Letters
Ghosal S
(2024)
A Unifying Framework for the Capacitated Vehicle Routing Problem Under Risk and Ambiguity
in Operations Research
Ho C
(2022)
Robust Phi-Divergence MDPs
Ho C.P.
(2022)
Robust ?-Divergence MDPs
in Advances in Neural Information Processing Systems
| Description | The number of phases (vapour, liquid or solid) present in a mixture at equilibrium under given conditions of temperature, pressure and global composition is a key consideration in the design of manufacturing processes and chemical products. The unexpected appearance of a new phase can impair the performance of a process or product, lead to wasted economic and material resources, or even call for a complete redesign. One of ADOPT's main use cases has been on computer-aided molecular and process design (CAMPD), where the integration of machine learning (ML) and mathematical optimisation was shown to significantly reduce the likelihood of designing unstable solvent mixtures, such as undergoing liquid-liquid demixing. This capability was first developed for a preset number of a few dozen candidate mixtures, used to train classifier surrogates for fast prediction of phase stability under different mixture compositions and temperatures. This work then led to the first classification approach capable of reliable quantitative predictions of a mixture's stability based on group contribution for a much larger set of candidate mixtures. |
| Exploitation Route | The ability to design and optimise solvent mixtures tailored to particular applications is relevant to a wide range of industries, including chemicals, pharmaceuticals, consumer goods, as well as carbon capture and storage (CCS). The findings of the ADOPT project are regularly shared with industrial partners through the Sargent Centre's industrial consortium and other industry partnerships through Imperial and UCL. |
| Sectors | Chemicals Energy Healthcare Pharmaceuticals and Medical Biotechnology |
| Title | Robust Hydrogen Infrastructure Model |
| Description | Robust Hydrogen Infrastructure Model and python code for generating robust representative days with uncertainty quantification |
| Type Of Material | Computer model/algorithm |
| Year Produced | 2024 |
| Provided To Others? | Yes |
| Impact | In this work, we employ data-driven robust optimisation to develop a framework for uncertainty-aware representative days explicitly characterised by polyhedral uncertainty sets. The proposed framework is applied on a multi-period mixed-integer linear model with dual temporal resolution which aims to determine the optimal yearly investment decisions and hourly operational decisions for the hydrogen infrastructure planning under demand uncertainty. To efficiently solve the large-scale two-stage adaptive robust optimisation problem, a hybrid decomposition algorithm is developed based on a two-step hierarchical procedure and the column-and-constraint generation method, which can significantly reduce the computational complexity. |
| URL | https://github.com/PRYSM-Lab/DataDrivenRobustEnergy |
| Description | Collaboration with IMT Lucca |
| Organisation | IMT Institute for Advanced Studies Lucca |
| Country | Italy |
| Sector | Academic/University |
| PI Contribution | We are co-developing a new set arithmetic for propagating tight bounds on expression trees in global optimisation. |
| Collaborator Contribution | Our partner at IMT Lucca apply this bounding capability in robust model-predictive control (MPC) of mechanical systems, such as obstacle avoidance |
| Impact | We are preparing a journal publication (for Journal of Global Optimisation or similar) |
| Start Year | 2024 |
| Description | Collaboration with TU Eindhoven |
| Organisation | Eindhoven University of Technology |
| Country | Netherlands |
| Sector | Academic/University |
| PI Contribution | We developed an approach for tightening the linear programming approximation of a graph neural network by shrinking the bounds in a branch&bound tree. |
| Collaborator Contribution | Our partners extended the open-source solver SCIP to accommodate these algorithms. |
| Impact | We have submitted a conference paper, this work is interdisciplinary between computer science (Imperial) and mathematics (TU Eindhoven). |
| Start Year | 2023 |
| Description | Collaboration with University of Connecticut & University of Wisconsin Madison |
| Organisation | University of Connecticut |
| Country | United States |
| Sector | Academic/University |
| PI Contribution | Our team contributes in the aspects of application development and focuses on the extension of the DOMINO framework to uncertainty considerations in a computationally efficient manner. |
| Collaborator Contribution | Dr Beykal and Dr Avraamidou have developed a framework for multilevel data-driven MINLP optimisation (DOMINO). |
| Impact | This collaboration focuses on data-driven multilevel MINLP optimisation with application to enterprise wide optimisation problems related to process planning, scheduling and pooling of chemical industries. We have presented results in the following conferences: AIChE Annual Meeting 2023, Foundations of Process/product Analytics and Machine learning (FOPAM) 2023. There is a forthcoming accepted conference paper in ESCAPE34 and we are currently in process of submitting a full-length article. |
| Start Year | 2023 |
| Description | Collaboration with University of Connecticut & University of Wisconsin Madison |
| Organisation | University of Wisconsin-Madison |
| Country | United States |
| Sector | Academic/University |
| PI Contribution | Our team contributes in the aspects of application development and focuses on the extension of the DOMINO framework to uncertainty considerations in a computationally efficient manner. |
| Collaborator Contribution | Dr Beykal and Dr Avraamidou have developed a framework for multilevel data-driven MINLP optimisation (DOMINO). |
| Impact | This collaboration focuses on data-driven multilevel MINLP optimisation with application to enterprise wide optimisation problems related to process planning, scheduling and pooling of chemical industries. We have presented results in the following conferences: AIChE Annual Meeting 2023, Foundations of Process/product Analytics and Machine learning (FOPAM) 2023. There is a forthcoming accepted conference paper in ESCAPE34 and we are currently in process of submitting a full-length article. |
| Start Year | 2023 |
| Description | KU Leuven |
| Organisation | University of Leuven |
| Department | VIB-KU Leuven Center for Cancer Biology |
| Country | Belgium |
| Sector | Public |
| PI Contribution | Dr Bongartz started as an Assistant Professor at KU Leuven in Sep-2022. The collaboration with Prof Chachuat and Prof Pantelides with Dr Bongartz on automating reduced-space formulations in global optimization continues. An implementation of the approach is being developed in the library MC++ developed in the group of Prof Chachuat. The work will be presented by Dr Bongartz at the next ESCAPE 33 conference to be hold in Athens, Greece in June 2023. |
| Collaborator Contribution | Dr Bongartz started as an Assistant Professor at KU Leuven in Sep-2022. The collaboration with Prof Chachuat and Prof Pantelides with Dr Bongartz on automating reduced-space formulations in global optimization continues. An implementation of the approach is being developed in the library MC++ developed in the group of Prof Chachuat. The work will be presented by Dr Bongartz at the ESCAPE 33 conference (https://escape33-ath.gr/) to be hold in Athens, Greece in June 2023. |
| Impact | - Conference paper to be presented at ESCAPE 33 conference (https://escape33-ath.gr/) to be hold in Athens, Greece in June 2023. - Open-source implementation of reduced-space formulation to be available in next release (2.2) of MC++ library (https://github.com/omega-icl/mcpp) |
| Start Year | 2022 |
| Title | CANON: Complete search Algorithms for Nonlinear OptimizatioN |
| Description | CANON is a C++ library for global optimization of nonlinear models using complete search algorithms. The complete-search algorithms current implemented in CANON are: (i) Reformulation of the MINLP model as an equivalent mixed-integer quadratically-constrained program (MIQCP), followed by MIQCP global optimization using the solver GUROBI. Various preprocessing strategies can be used, including local search and bounds tightening, prior to the MIQCP reformulation. The exact reformulation of polynomial subexpressions into quadratic form relies on the introduction of auxiliary variables and constraints. A number of nonlinear terms cam also be handled by GUROBI via their approximation as piecewise linear expressions; or, alternatively, they can be enclosed by polynomial models within CANON. (ii) Piecewise linearization of the MINLP model by progressively adding break-points to a MIP relaxation, thereby creating a convergent sequence of relaxations. The MIP relaxations are solved using GUROBI. Various preprocessing strategies can be used, including local search and bounds tightening, prior to the MIP relaxation hierarchy. |
| Type Of Technology | Software |
| Year Produced | 2022 |
| Open Source License? | Yes |
| Impact | The quadratic reformulation approach and performance were highlighted in a talk by Gurobi CEO, Ed Rothberg during AIChE Annual Meeting 2023. |
| Title | CRONOS - Complete seaRch sOlutions for NOnlinear Systems |
| Description | CRONOS provides a collection of C++ classes for evaluation and uncertainty propagation in implicit nonlinear systems, including the solutions of nonlinear algebraic and differential equations. It builds on the DAG construction capability of MC++ in their evaluation in diverse set arithmetics for the uncertaitny propagation. |
| Type Of Technology | Software |
| Year Produced | 2022 |
| Open Source License? | Yes |
| Impact | CRONOS is used by the CANON (local and global optimisation of continuous and mixed-integer models) and MAGNUS (model development and analysis under uncertainty) to handle problems with dynamic systems embedded (e.g. dynamic parameter estimation, optimal control). |
| Title | Graph neural network capability in OMLT |
| Description | OMLT is an open-source tool that translates from a machine learning representation into an optimisation representation. We use OMLT, for instance in collaborations with BASF and the Imperial Business School, to solve inverse problems over already-trained machine learning models. The new component of OMLT, which was developed as part of ADOPT, is to solve inverse problems over graph neural networks. |
| Type Of Technology | Software |
| Year Produced | 2024 |
| Open Source License? | Yes |
| Impact | We have developed an approach, in collaboration with BASF, to suggest possible new molecules for development. |
| URL | https://github.com/cog-imperial/OMLT/pull/110 |
| Title | MC++ Toolkit for Construction, Manipulation and Bounding of Factorable Functions |
| Description | MC++ (currently version 4.0) provides a collection of C++ classes for construction, manipulation and evaluation of factorable functions using directed acyclic graphs (DAGs). The use of DAGs enables manipulating expressions, including automatic differentiation, lifting of polynomial subexpressions, quadratization of polynomial subexpressions, elimination of auxiliary/intermediate variables, and identification of reduction constraints. MC++ also supports the inclusion of external operations, which enables implicit functions such as differential equations, neural networks or even optimisation subproblems to be included in a DAG. The evaluation methods in MC++ include a range set arithmetics, namely interval bounds, spectral bounds, convex/concave relaxations, Taylor and Chebyshev model estimators, polyhedral relaxations, and superposition models. Since version 4.0, a Python interface called pyMC, is being actively developed to expose the DAG construction, manipulation and evaluation capability. |
| Type Of Technology | Software |
| Year Produced | 2024 |
| Open Source License? | Yes |
| Impact | C++ provides the basis for evaluating McCormick relaxations in the global solver MAiNGO developed by the group of Professor Mitsos at RWTH Aachen (https://avt-svt.pages.rwth-aachen.de/public/maingo/). It also provides the basis for the libraries CRONOS (evaluation and relaxation of implicit models described by algebraic and/or differential equations), CANON (local and global optimisation of continuous and mixed-integer models), and MAGNUS (model development and analysis under uncertainty) developed by the Chachuat group at Imperial College London. |
| Description | ADOPT Technical Day 2022 |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Technical day organized between the UK team (Imperial, UCL), German team (RWTH Aachen), and industry partners (Gurobi, Siemens PSE). The following keynote talks were presented during this event: - Optimization challenges in CAMPD (Prof Adjiman) - Challenges in short-term planning of IRPC (Prof Chachuat) as well as research updates on: - Global optimization of polynomial programs via quadratization (Tanuj Karia) - Automatic construction of reduced-space global optimization formulation (Dr Bongartz) All the talks were followed by discussions. The meeting took place in hybrid mode. |
| Year(s) Of Engagement Activity | 2022 |
| Description | ADOPT Technical Day 2023 |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Technical day organized between the UK team (Imperial, UCL), German team (RWTH Aachen), and industry partners (Gurobi, Siemens PSE). The following keynote talks were presented during this event: - Challenges in renewable supply chain design (Prof Grit Walther & Simon Thomä) - GNN-MIP-CAMD: Computer-aided molecular design based on mixed-integer programming on graph neural network (Shiqiang Zhang) - Subdomain separability in global optimization (Dr Jens Deussen) - Interval superposition models for global optimization (Yanlin Zha) All the talks were followed by discussions. The meeting took place in hybrid mode. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Dr Bongartz Seminar |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | National |
| Primary Audience | Postgraduate students |
| Results and Impact | Dr Bongartz gave a seminar to the Sargent Centre community, entitled "Accelerating global optimization for process design: Reduced-space formulations, relaxations & parallelization". The seminar was delivered in hybrid mode to reach out to more students and researchers. It triggered a collaboration with PhD student Tanuj Karia on reduced-space formulations for global optimization and led to a 4-month secondment of Tanuj Karia at RWTH Aachen between Oct-2022 - Jan-2023. |
| Year(s) Of Engagement Activity | 2022 |
| Description | School visit (London) |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Local |
| Primary Audience | Schools |
| Results and Impact | 40 pupils attended a talk on chemical engineering in which research topics were presented. |
| Year(s) Of Engagement Activity | 2025 |