GALINI: Global ALgorithms for mixed-Integer Nonlinear optimisation of Industrial systems

Lead Research Organisation: Imperial College London
Department Name: Dept of Computing

Abstract

At the 2015 Paris climate conference, 195 countries agreed that global greenhouse gases should peak as soon as possible and that countries should thereafter rapidly reduce their emissions. The process industries must therefore reduce their energy consumption and increase efficiency while maintaining consumer services. Next generation decision-making software at the interface of engineering, computer science, and mathematics is critical for these efficient systems of the future. Already, state-of-the-art computational packages are routine in the process industries; practically every major company uses simulation and optimisation to model production in different modes including: continuous, batch, and semi-continuous production systems. But more efficient industrial systems require simultaneously considering many tightly integrated subsystems which exponentially increase complexity and necessitate many temporal/spatial scales; the resulting decision making problems may not be solvable with current techniques. Increasing efficiency may also jeopardise safety: the process integration required for efficiency implies interchanging heat between processes and may damage safety precautions by transferring disturbances across a plant.

During this fellowship, we propose to develop GALINI, new decision-making software constructing and deploying next generation process optimisation tools dealing with combinatorial complexity, disparate temporal/spatial scales, and safety considerations. The GALINI project proposes step-changes in optimisation algorithms that are immediately applicable to efficiency challenges in process systems engineering (PSE): safely operating batch reactors, retrofitting heat-exchanger networks, intermediate blending, and integrating planning and scheduling. We will freely release our software on open-source platform Pyomo and build an international user community.

The primary GALINI research aim is to develop optimisation software that pushes the boundary of computational tractability for PSE energy efficiency applications. Effective optimisation software in the process industries answers: How can we best achieve a definite engineering objective? Given constraints such as an existing plant layout or a contractual obligation to produce specific products, the software supports novel engineering by quantitatively comparing the implications of different options and identifying the best decision. GALINI is particularly interested in design: How should we build new facilities or modify existing ones to achieve our design goals with maximum efficiency?

The state-of-the-art in decision making for the process industries is represented by commercial modelling software such as AspenTech and gPROMS. Practically every major company in the process industries uses these software tools since the outputs of the simulation or optimisation can be implemented with minimal day-to-day operational disruption and savings can be realised with a payback time as short as 6-12 months. GALINI will develop deterministic global optimisation software for mixed-integer nonlinear programs, a type of optimisation problem highly relevant to energy efficiency and process systems engineering. Energy efficiency instances may exhibit the mathematical property of nonconvexity, i.e. have many locally optimal solutions; global optimisation mathematically guarantees the best process engineering solution. GALINI proposes transformational shifts in algorithms that creatively reimagine the core divide-and-conquer algorithm typically applied to this type of optimisation problem. Our approach is to freely release GALINI to users including those in the process industries, publicise the software, demonstrate its utility, and build a user community that will feed back into software development.

Planned Impact

The GALINI fellowship will positively impact a wide variety of groups beyond the academy, both in the UK and internationally. Here we summarise the impact on: the process industries, solver software developers, policy makers, schools, and the wider public.

Process Industry:

+ The novel engineering research into retrofitting heat exchangers, integrating planning and scheduling, and safely operating batch reactors have immediate, domain specific benefits for the process industries;
+ The advanced decision making tool GALINI can better help the process industries increase energy efficiency and profitability;
+ We are partnering with several software vendors working with the process industries, so advances will be available either through our own GALINI code base or through parts of the research that the project partners chose to pick-up and distribute in their own venues.

Solver Software Developers:

+ We are making our code open-source, so all or part of it may feed into future open-source or commercial optimisation software;
+ We are developing fundamental new algorithms that will drive the state-of-the-art and will encourage more companies to move into this domain.

Policy & Other Decision Makers:

+ Our code will be freely available to policy and decision makers from around the world;
+ GALINI will be able to ask the question "What is it that we need to guarantee the objective that we want?" and will be able to identify gaps where, for example, no existing technology cannot reach a certain engineering objective. This may help guide polices;
+ GALINI will also be able to identify when an energy efficiency policy should be possible for engineering, this could help policy makers choose regulatory structures.

Schools & Wider Public:

+ We will attract advanced undergraduates to our work by allowing them to participate in the user workshops that we will run throughout the UK;
+ We will communicate our work to the general public via events such as Royal Society Summer Science Exhibition and Imperial Festival.

The primary activities for the proposed GALINI fellowship are research software development, novel algorithms, and new engineering research. But we will also release GALINI to process industries users, publicise the software, demonstrate its utility, build a user community that will feed back into software development, and communicate our work with the general public. The Pathways to Impact document explains how we will accomplish these plans.

Publications

10 25 50
publication icon
Campos J (2019) A multilevel analysis of the Lasserre hierarchy in European Journal of Operational Research

publication icon
Ceccon F (2019) SUSPECT: MINLP special structure detector for Pyomo in Optimization Letters

publication icon
Mistry M (2018) Satisfiability modulo theories for process systems engineering in Computers & Chemical Engineering

publication icon
Olofsson S (2019) GPdoemd: A Python package for design of experiments for model discrimination in Computers & Chemical Engineering

publication icon
Wiebe J (2018) Data-Driven Optimization of Processes with Degrading Equipment in Industrial & Engineering Chemistry Research

publication icon
Wiebe J (2019) Robust Optimization for the Pooling Problem in Industrial & Engineering Chemistry Research

 
Description Only 1 of the 5 years of the grant has passed, but we are well on track to achieve the aims of the grant. We have developed GPdoemd, a Python package for design of experiments for model discrimination, and SUSPECT, a Python package for detecting special mathematical structure in optimisation problems. Both of these packages are useful on their own, but they're also key developments for executing the remaining open source software to be developed on the grant.

We have also discovered and developed new methodology for solving difficult optimisation problems, e.g. heuristics with performance guarantees and cutting planes. We're working on integrating this new methodology into the larger software package that will be developed as a part of this grant.
Exploitation Route The available open source software can be used by researchers world-wide for free. This open source software may have implications for energy and healthcare sectors.
Sectors Energy,Healthcare

URL https://github.com/cog-imperial
 
Description The findings from our research have been taken up by the Royal Mail Data Science Group, which is evaluating our contributions towards minimising the number of delivery vans they need. Our findings are also being evaluated by BASF and Schlumberger. Our findings have presented to MPs and policy makers in Parliament as a part of the 2019 STEM for Britain competition. Our open source software packages GPdoemd and SUSPECT are being used for further research by other groups.
First Year Of Impact 2018
Sector Energy,Healthcare
Impact Types Economic,Policy & public services

 
Description BASF PhD Studentship
Amount £270,000 (GBP)
Organisation BASF 
Sector Private
Country Germany
Start 06/2019 
End 05/2022
 
Description Industrial CASE Account - Imperial College London 2017
Amount £1,499,328 (GBP)
Funding ID EP/R511961/1 
Organisation Engineering and Physical Sciences Research Council (EPSRC) 
Sector Academic/University
Country United Kingdom
Start 09/2017 
End 09/2022
 
Description Newton International Fellowship to Dr Jan Kronqvist
Amount £99,000 (GBP)
Organisation The Royal Society 
Sector Academic/University
Country United Kingdom
Start 03/2019 
End 02/2021
 
Title Test set for the minimum number of matches problem 
Description Heat exchanger network synthesis is a critically important engineering problem improving energy recovery in chemical processes. The minimum number of matches problem is a difficult optimisation problem and a bottleneck for designing good heat recovery networks. Many authors have discussed this optimisation problem over the years, but we have collected and made publicly available this test set for the first time. Now researchers will have access to all 51 problems in the open literature. 
Type Of Material Database/Collection of data 
Year Produced 2017 
Provided To Others? Yes  
Impact We are enabling follow-on research because many of these test instances had been described in an academic paper but never made easily-accessible. By providing researchers both with (i) the complete set of test problems in the open literature and (ii) our methods for finding feasible solutions, we make it possible for future research in this area to more properly metric use itself. 
URL https://github.com/cog-imperial/min_matches_heuristics/tree/master/data/original_instances
 
Description BASF - Statistics and Machine Learning for Chemicals 
Organisation BASF
Country Germany 
Sector Private 
PI Contribution Using a problem relevant to BASF, my team investigates a large-scale industrially-relevant optimisation problem that includes machine-learning tree-based models in the objective function. Our methodology leverages decomposition structure in the optimisation problem. Our work directly impacts BASF's energy efficiency and is easily extended to other applications, e.g. materials design and finance, since tree-based models are well suited for modelling unknown nonlinear functions. Furthermore, our optimisation problem may broadly incorporate other machine-learning models, contributing to the design of a unifying framework where mathematical optimisation is integrated with machine-learning and data analytics for effective decision making. My team contributed the expertise and developed the methodology.
Collaborator Contribution BASF finds machine-learning tree-based models effective for modelling catalyst behaviour because closed-form mathematical expressions are not known for many chemical catalysis applications. Catalysts are essential for energy efficiency at BASF. Therefore, developing the best-performing catalysts requires optimising over their tree-based models. BASF proposed a large-scale industrially-relevant optimisation problem that contains machine-learning tree-based models in the objective function. The optimisation problem considers the tree-based models but also penalises solutions that are far from training data, i.e. it optimises the tree-based models closer to the experimental data. BASF also contributed £40k in initial seed funding to fund my PhD student associated with this project.
Impact Based on the outcomes of this project, BASF agreed to fund a full PhD scholarship (£270k). This PhD student will be associated with the EPSRC projects and partially trained by researchers associated with the EPSRC projects. PhD student Miten Mistry (paid by BASF) collaborated with PDRA Dimitris Letsios and PI Ruth Misener to develop a submission to the 2019 STEM for Britain competition. Miten's accomplishment is documented here: http://www.imperial.ac.uk/news/190330/department-computing-researchers-selected-present-research/
Start Year 2017
 
Description Bayer - Design of Experiments for Model Discrimination 
Organisation Bayer
Country Germany 
Sector Private 
PI Contribution My team developed the methodology and an open source software package GPdoemd (https://github.com/cog-imperial/GPdoemd) for design of experiments for model discrimination. The theoretical contribution to design of experiments for model discrimination is to replace the original models with Gaussian process surrogate models. By using Gaussian process surrogates, we make the entire model discrimination process more accessible.
Collaborator Contribution Bayer contributed their time and developed case studies relevant to their business.
Impact We developed open source software GPdoemd (https://github.com/cog-imperial/GPdoemd) and released in publicly. Several research groups worldwide are trialling the software. PhD student Simon Olofsson (not funded by the EPSRC, but collaborating with Dr Ruth Misener who is funded by the EPSRC) presented his research at the 2019 STEM for Britain competition (http://www.imperial.ac.uk/news/190330/department-computing-researchers-selected-present-research/). We have published a joint Imperial/Bayer paper (DOI 10.1016/j.compchemeng.2019.03.010).
Start Year 2017
 
Description Collaboration with the research group of Processor Liesbet Geris 
Organisation University of Liege
Country Belgium 
Sector Academic/University 
PI Contribution We provided the expertise in optimisation under uncertainty.
Collaborator Contribution The group of Professor Liesbet Geris provided the application and the computing resources. PhD student Mohammad Mehrian visited the group of Ruth Misener to collaborate on this project. PhD student Simon Olofsson (not funded by the EPSRC, but Simon's training leads him to collaborate on the project) also visited the group of Professor Geris.
Impact We have published two full-length journal papers with Professor Geris' group: one is in Biotechnology & Bioengineering (2018) and the other is in IEEE Transactions on Biomedical Engineering (2019). We have also published several conference papers together.
Start Year 2016
 
Description Royal Mail Data Science Group 
Organisation Royal Mail plc
PI Contribution The Royal Mail wants to minimise the number of vans required at each of their "Delivery Offices" or DOs. Royal Mail has approximately 1250 Delivery Offices and 43k vans, so reducing the number of vans is a big cost (and environmental) savings for the company. My research team has developed a methodology for responding quickly to uncertainty. In our work, we design a schedule the night before based on historical data. Then, when anything goes wrong on the day, our recovery strategy quickly adapts to the change. We show that optimising the night-before schedule in a good way, i.e. with lexicographic optimisation, helps recovery on the day itself. Moreover, our results make Royal Mail more comfortable that they can reduce the number of vans without penalty.
Collaborator Contribution Our collaborators in the Royal Mail Data Science Group have meet with us for ~5 full days over the course of the last 6 months. Our collaborators also have provided us with test sets to try out our ideas.
Impact Royal Mail has offered to fund an EPSRC CASE conversion PhD studentship for MSc student Ms Natasha Page who collaborated on the project (Imperial is holding its internal competition for these studentships in mid-March). Royal Mail has also offered to write a letter supporting one of the REF Impact Case studies. We are working to prepare a joint paper.
Start Year 2018
 
Description Sandia National Laboratories 
Organisation Sandia Laboratories
Country United States 
Sector Private 
PI Contribution We present SUSPECT, an open source toolkit that symbolically analyses mixed-integer nonlinear optimisation problems formulated using the Python algebraic modelling library Pyomo. This SUSPECT toolkit is the first step in releasing the open-source software GALINI promised in the grant.
Collaborator Contribution Our contact at Sandia, Dr John Siirola, develops Pyomo, a Python-based Algebraic Modelling Language for optimisation. Dr Siirola helped us integrate SUSPECT with Pyomo.
Impact We presented SUSPECT to potential users at the 2019 Centre for Process Systems Engineering Annual Industrial Consortium Meeting (~20 potential users). We have published a joint Imperial/Sandia paper (https://link.springer.com/article/10.1007/s11590-019-01396-y).
Start Year 2017
 
Description Schlumberger 
Organisation Schlumberger Limited
Department Schlumberger Cambridge Research
Country United Kingdom 
Sector Academic/University 
PI Contribution My team has been developing methodologies for optimisation under uncertainty.
Collaborator Contribution Schlumberger has contributed monies for an EPSRC iCASE studentship and worked jointly on the research together.
Impact Pending successful translation of our ongoing research, Schlumberger has offered to write a letter for a potential REF Impact Case Study. PhD student Johannes Wiebe (supported by the PhD student from Schlumberger, but collaborating with PI Ruth Misener whose time is supported by the EPSRC) will be seconded at Schlumberger Research Cambridge for 2 months in 2019. We have one joint Imperial/Schlumberger journal paper (https://pubs.acs.org/doi/abs/10.1021/acs.iecr.8b03292) and 2 conference papers under submission.
Start Year 2017
 
Title GPdoemd 
Description GPdoemd is a Python package for design of experiments for model discrimination using Gaussian process surrogate models. 
Type Of Technology Software 
Year Produced 2018 
Open Source License? Yes  
Impact Simon Olofsson presented his software package at the 2019 Centre for Process Systems Engineering Annual Industrial Meeting. Simon Olofsson presented his research at the 2019 STEM for Britain competition. 
URL http://www.imperial.ac.uk/news/190330/department-computing-researchers-selected-present-research/
 
Title Heuristics with Performance Guarantees for the Minimum Number of Matches Problem in Heat Recovery Network Design 
Description This is an implementation of our paper (10.1016/j.compchemeng.2018.03.002, Computers & Chemical Engineering). Heat exchanger network synthesis exploits excess heat by integrating process hot and cold streams and improves energy efficiency by reducing utility usage. Determining provably good solutions to the minimum number of matches is a bottleneck of designing a heat recovery network using the sequential method. This software develops heuristic methods with performance guarantees using three approaches: (i) relaxation rounding, (ii) water filling, and (iii) greedy packing. Also available in this Github repository are numerical results from a collection of 51 instances substantiate the strength of the methods. 
Type Of Technology Software 
Year Produced 2017 
Open Source License? Yes  
Impact Other researchers have tested our methods. 
URL https://github.com/cog-imperial/min_matches_heuristics
 
Title SUSPECT 
Description SUSPECT is an open source toolkit that symbolically analyses mixed-integer nonlinear optimisation problems formulated using the Python algebraic modelling library Pyomo. 
Type Of Technology Software 
Year Produced 2018 
Open Source License? Yes  
Impact Francesco Ceccon presented his work at the 2019 Centre for Process Systems Engineering Annual Industrial Consortium Meeting. 
 
Description CPSE Annual Industrial Consortium Meeting 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Professional Practitioners
Results and Impact My team presented their research in the Centre for Process Systems Engineering Annual Industrial Consortium meeting. There, we received feedback on our work and requests for further collaborations.
Year(s) Of Engagement Activity 2017,2018,2019
 
Description STEM for Britain 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Policymakers/politicians
Results and Impact Two of my PhD students participated in the 2019 STEM for Britain competition (http://www.setforbritain.org.uk/index.asp) where the presented my team's contributions to MPs from both Houses of Parliament.
Year(s) Of Engagement Activity 2019
URL http://www.imperial.ac.uk/news/190330/department-computing-researchers-selected-present-research/
 
Description Seminar at Carnegie Mellon 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Postgraduate students
Results and Impact 80 postgraduate students and their supervisors attended a seminar given by Ruth Misener, which led to discussions on the topic and interest in future research
Year(s) Of Engagement Activity 2019
URL https://twitter.com/crislopeslara/status/1102990243752542208
 
Description Speed mentoring for AnitaB.org at the Twitter London office 
Form Of Engagement Activity A formal working group, expert panel or dialogue
Part Of Official Scheme? No
Geographic Reach Local
Primary Audience Postgraduate students
Results and Impact I will be attending a speed mentoring event for AnitaB.org on Thursday, 15 March at the Twitter London office. This mentorship activity for young women in the computer science community to get mentorship from more senior members such as myself.
Year(s) Of Engagement Activity 2018
URL https://community.anitab.org/event/meet-your-local-mentors/