Département Gestion des eaux urbaines

EMUmore - EMUlation allows us to do MORE with our models

Did you ever feel that your computer model was too slow? Not only to finally produce some output you could look at, but also to analyse the sensitivity of the results to the many parameters you had to specify. And performing a full-fledged uncertainty analysis is beyond scope anyways, because for the many ten or hundred thousand model runs, your simulator is simply too time consuming? Yes, we hear you.

Emulators to the rescue

The direct consequence of time consuming simulators is that researchers work less efficiently and that results can be questionable because parameter values are chosen on an ad-hoc basis and not based on all the available evidence.

There are four approaches to deal with this problem: i) discard systems analysis and intelligent control strategies, ii) work only with utterly simplified models, iii) use high-performance computing (HPC), and iv) construct fast surrogate models. The first two are inefficient and the third, using HPC, often requires considerable investment in HPC know-how and IT equipment. It often also requires re-programming simulation software, which is not always the best strategy for all projects and problems. Therefore, constructing fast surrogate models, so-called “emulators”, to speed up slow simulators is very attractive. It does not require a huge investment in new hardware and software, and the same tool can be used to solve very different problems.

However, although we have obtained promising results for the emulation of receiving waters and urban drainage systems, we find that the emulation tools we have developed so far are not yet ready to tackle nonlinear problems. Also, it is unclear in how far they can improve the operation of urban water systems and how they can be used to fuse the information from wireless sensor networks, with potentially hundreds of sensors.

A complicated simulator (left) with many internal states is replaced by an emulator that quickly evaluates an input-output relation learned from data (right).

Goals

  1. Improve and develop emulation methods We will continue improving our Gaussian Processes based emulation methods and incorporate algorithms from Reduced Basis Methods.
  2. Demonstrate the usefulness of emulators for urban wastewater systems We will provide many realistic case studies where emulation give an edge in System Identification (a.k.a. model calibration), Design Optimization, and Model Predictive Control.
  3. Disseminate the capacity to construct emulators to practitioners During this project we will organize 3 dissemination events to make our tools accessible to everybody.
  4. Support Open Science by doing open science The outputs of the project will be released under Libre Software Licenses and Creative Common Licenses.

The current status

EMU more started in May 2017 and will run for 18 months.

Check our bitbucket project page for progress, case studies and issues. If you'd like to contribute, please get in touch.

An emu. A fast running bird. A fast running surrogate model. ... What a coincidence! (Yes, this is our idea of a logo. Yes, we know. If you have a better idea, please get in touch.)

News

Check out EMUmore's blog!

Follow us