Clustering, Labelling and Predicting: Making use of the ML toolbox for battery science

Transcription

1 1 Clustering, Labelling and Predicting: Making use of the ML toolbox for battery science Dr Sam Cooper March 19

2 Electrochemical Science and Engineering Lithium Ion Material Analysis Solid State Supercapacitors Microstructure In-operando diagnostics Understanding degradation Improving performance Isotopic methods Bulk diffusivities Surface exchange coefficients Grain boundary analysis Understanding new technology Modelling mechanical changes Thermal diagnostics Flexible batteries Thermal & physical models Supercapacitors, lithium hybrid capacitors, structural capacitors (body panels) Hybridisation (batteries/fc) Acquire 3D microstructural dataset using tomography Simulate transport processes at the micro-scale Battery Pack Design Modelling Nanofiber systems Techno-Economics Future chemistries Long Distances Heavy Vehicles Hydrogen & Fuel Cells Mass Market Plug-in-Hybrids Battery Electric Short Distances Light Vehicles Plug-in-Hybrids Advanced Biofuels Mass Market Note: Not to scale time Long Distances Heavy Vehicles Pack design BMS integration & control Thermal management Physics based models Empirical physics models Reduced order models Multiple applications Flexible cells Novel production routes No current collectors Orientation analysis Market size & trends Business opportunities Investment decisions Sustainability analysis Sodium solid-state Zinc-air batteries and dendrites Electrolyte additives/separators Silicon anode volume expansion 2

3 3 Outline Sorting second life cells Grid storage modelling and prediction Convolution for microstructure

4 Cell sorting with Differential Thermal Voltammetry 4

5 Introduction to DTV Differential thermal voltammetry DTV Galvanostatic operation (typically discharge) 5 B. Wu, V. Yufit, Y. Merla, R. F. Martinez-Botas, N. P. Brandon, and G. J. Offer, Differential thermal voltammetry for tracking of degradation in lithium-ion batteries, J. Power Sources, vol. 273, pp , Jan

6 Interpretation of DTV 6 Y. Merla, B. Wu, V. Yufit, N. P. Brandon, R. F. Martinez-Botas, and G. J. Offer, Extending battery life: A low-cost practical diagnostic technique for lithium-ion batteries, J. Power Sources, vol. 331, pp , Nov

7 7 Application to second-life Can we put second life automotive cells into lower spec. devices? Innovate UK project with M- POA solar and Denchi Power DTV offers a metric of similarity Better cell matching to extending pack life

8 8 Variation

9 Capacity / Ah Similar cell clustering - DTV Differential thermal voltammetry used to characterise second life cells. Wu, B. et al. Differential thermal voltammetry for tracking of degradation in lithium-ion batteries. J. Power Sources 273, (2015). High dimensional clustering algorithm groups similar cell to maximise pack life. 9 Internal resistance / Ω

10 Reinforcement Learning for Grid-scale storage 10

11 11 Grid-scale storage Opportunity for batteries: Enhanced frequency response GreenHedge London based company

12 Battery SoH data 12 J.-Y. Morille - Degradation of lithium-ion batteries, MSc Thesis, Imperial College London (2017)

13 13 Fitting with Piantadosi, S. T. One parameter is always enough.

14 Scenario testing simple models 14 D. Reece, Modelling the Hybridisation of Grid-Level Energy Storage for Frequency Response, MSc Thesis, Imperial College London (2018)

15 15 Typical performance is not the whole problem

16 16 Frequency data + prediction

17 17 Motivation for enhanced control

18 Microstructural analysis with convolutional networks

19 19 Electrodes Coupled, multi-phase, multi-physics systems Performance controlled by many parameters Parameters depend on state and evolve over time

20 FIB-SIMS FIB-SEM X-ray μm μm μm Cooper, S. J. et al. Multi Length-Scale Quantification of Hierarchical Microstructure in Designed Microtubular SOFC Electrodes. in ECS Transactions 68, (Ecs, 2015). 20

21 Monster soup 21

22 22??

23 23 Segmentation - conventional Manual segmentation A nightmare Thresholding Usually inappropriate for 2 phase systems Always inappropriate for 3 phase systems Algorithmic E.g. watershedding rarely aligns with perception and struggles with artefacts

24 24 Segmentation ML style Supervised learning for 3D image classification. Spatial convolution gives each voxel context and far outperforms conventional methods.

25 25 Compression & Synthesis

26 26 Segmented data then what? ML

27 27 Designed electrodes Hierarchical electrodes fabricated at Oxford Nano-scale porosity/nanotube model - homogenised using TauFactor 2D model built by Antonio Bertei NP + Diffusion + BV

28 28 Evolutionary optimisation Adjust the latent space parameters to navigate the space of possible microstructures.

29 29 Adversarial warning! Optimisation may just find your assumptions The more complex your model is, the harder this can be to spot

30 Microstructure optimisation Structure electrode by Prof Grant and Dr Huang at Oxford 30 C. Huang, N. P. Young, J. Zhang, H. J. Snaith, and P. S. Grant, A two layer electrode structure for improved Li Ion diffusion and volumetric capacity in Li Ion batteries, Nano Energy, vol. 31, pp , 2017.

31 31

32 32 Mathematics for machine learning

33 33 Interactive widgets

34 Thanks to the ESE group 34 Academics/Mgmt. Team Dr Billy Wu Dr Chandramohan George Dr Gregory Offer Dr Huizhi Wang Dr Jacqueline Edge Dr Monica Marinescu Prof. Nigel Brandon Dr Sam Cooper Dr Yatish Patel Nicholas Dean PDRAs (20) Dr Aayan Banerjee Dr Alastair Hales Dr Alex Holland Dr Binbin Chen Dr Chenzhen Ji Dr Ganesh Madabattula Dr Jingyi Chen Dr Jorge Barreras Dr Karthik Radhakrishnan Dr Tasnuva Khaleque Dr Laura Bravo Diaz Dr Natasha Fiig Dr Nina Meddings Dr Shen Li Dr Simon O Kane Dr Teddy Szemberg O Connor Waseem Marzook Dr Weilong Ai Dr Xinhua Liu Dr Youxiu Wei PhD students (17) Andrea Gayon Lombardo Anna Tomaszewska Ashkan Kavei Edouard Querel Emma Vendola Ian Campbell Khairul Bin Muhammed Krishna Gopalakrishnan Max Naylor Marlow Mei Chin Pang Oisin Shaw Ridwanur Chowdury Ryan Prosser Tianyu Zhao Weiqiang Tang Xiao Hua Yan Zhao Recruiting: PM, Tech, 8-10x PDRAs, xphds