Theory-Guided Data Science, A Petrophysical Case Study from the Diyab Formation

A practical example of a theory-guided data science case study is presented to evaluate the potential of the Diyab formation, an Upper Jurassic interval, source rock of some of the largest reservoirs in the Arabian Peninsula. A workflow base on a three-step approach combining the physics of logging tool response and a probabilistic machine-learning algorithm was undertaken to evaluate four wells of the prospect. At first, a core-calibrated multi-mineral model was established on a concept well for which an extensive suite of logs and core measurements had been acquired. To transfer the knowledge gained from the latter physics-driven interpretation onto the other data-scarce wells, the relationship between the output rock and fluid volumes and their input log responses was then learned by means of a Gaussian Process Regression (GPR). Finally, once trained on the key well, the latter probabilistic algorithm was deployed on the three remaining wells to predict reservoir properties, quantify resource potential and estimate volumetric-related uncertainties. The physics-informed machine-learning approach introduced in this work was found to provide results which matches with the majority of the available core data, while discrepancies could generally be explained by the occurrence of laminations which thickness are under the resolution of nuclear logs. Overall, the GPR approach seems to enable an efficient transfer of knowledge from data-rich key wells to other data-scarce wells. As opposed to a more conventional formation evaluation process which is carried out more independently from the key well, the present approach ensures that the final petrophysical interpretation reflects and benefits from the insights and the physics-driven coherency achieved at key well location.

Impedance-based forecasting of battery performance amid uneven usage

Accurate forecasting of lithium-ion battery performance is important for easing consumer concerns about the safety and reliability of electric vehicles. Most research on battery health prognostics focuses on the R&D setting where cells are subjected to the same usage patterns, yet in practice there is great variability in use across cells and cycles, making forecasting much more challenging. Here, we address this challenge by combining electrochemical impedance spectroscopy (EIS), a non-invasive measurement of battery state, with probabilistic machine learning. We generated a dataset of 40 commercial lithium-ion coin cells cycled under multistage constant current charging/discharging, with currents randomly changed between cycles to emulate realistic use patterns. We show that future discharge capacities can be predicted with calibrated uncertainties, given the future cycling protocol and a single EIS measurement made just before charging, and without any knowledge of usage history. Our method is data-efficient, requiring just eight cells to achieve a test error of less than 10%, and robust to dataset shifts. Our model can forecast well into the future, attaining a test error of less than 10% when projecting 32 cycles ahead. Further, we find that model performance can be boosted by 25% by augmenting EIS with additional features derived from historical capacity-voltage curves. Our results suggest that battery health is better quantified by a multidimensional vector rather than a scalar State of Health, thus deriving informative electrochemical `biomarkers' in tandem with machine learning is key to predictive battery management and control.

Winsorization for Robust Bayesian Neural Networks

With the advent of big data and the popularity of black-box deep learning methods, it is imperative to address the robustness of neural networks to noise and outliers. We propose the use of Winsorization to recover model performances when the data may have outliers and other aberrant observations. We provide a comparative analysis of several probabilistic artificial intelligence and machine learning techniques for supervised learning case studies. Broadly, Winsorization is a versatile technique for accounting for outliers in data. However, different probabilistic machine learning techniques have different levels of efficiency when used on outlier-prone data, with or without Winsorization. We notice that Gaussian processes are extremely vulnerable to outliers, while deep learning techniques in general are more robust.

SPUCL (Scientific Publication Classifier): A Human-Readable Labelling System for Scientific Publications

To assess critically the scientific literature is a very challenging task; in general it requires analysing a lot of documents to define the state-of-the-art of a research field and classifying them. The documents classifier systems have tried to address this problem by different techniques such as probabilistic, machine learning and neural networks models. One of the most popular document classification approaches is the LDA (Latent Dirichlet Allocation), a probabilistic topic model. One of the main issues of the LDA approach is that the retrieved topics are a collection of terms with their probabilities and it does not have a human-readable form. This paper defines an approach to make LDA topics comprehensible for humans by the exploitation of the Word2Vec approach.

Uncertainty Quantification and Bayesian Active Learning for Rupture Life Prediction in Ferritic-Martensitic Steels

Three probabilistic methodologies are developed for predicting the long-term creep rupture life of 9−12 𝑤𝑡% 𝐶𝑟 ferritic-martensitic steels using their chemical and processing parameters. The framework developed in this research strives to simultaneously make efficient inference along with associated risk, i.e., the uncertainty of estimation. The study highlights the limitations of applying probabilistic machine learning to model creep life and provides suggestions as to how this might be alleviated to make an efficient and accurate model with the evaluation of epistemic uncertainty of each prediction. Based on extensive experimentation, Gaussian Process Regression yielded more accurate inference (𝑃𝑒𝑎𝑟𝑠𝑜𝑛 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛 𝑐𝑜𝑒𝑓𝑓𝑖𝑐𝑒𝑛𝑡> 0.95 for the holdout test set) in addition to meaningful uncertainty estimate (i.e., coverage ranges from 94 – 98% for the test set) as compared to quantile regression and natural gradient boosting algorithm. Furthermore, the possibility of an active learning framework to iteratively explore the material space intelligently was demonstrated by simulating the experimental data collection process. This framework can be subsequently deployed to improve model performance or to explore new alloy domains with minimal experimental effort.