Molecular property calculations are the bedrock of chemical physics. High-fidelity \textit{ab initio} modeling techniques for computing the molecular properties can be prohibitively expensive, and motivate the development of machine-learning models that make the same predictions more efficiently. Training graph neural networks over large molecular databases introduces unique computational challenges such as the need to process millions of small graphs with variable size and support communication patterns that are distinct from learning over large graphs such as social networks. This paper demonstrates a novel hardware-software co-design approach to scale up the training of graph neural networks for molecular property prediction. We introduce an algorithm to coalesce the batches of molecular graphs into fixed size packs to eliminate redundant computation and memory associated with alternative padding techniques and improve throughput via minimizing communication. We demonstrate the effectiveness of our co-design approach by providing an implementation of a well-established molecular property prediction model on the Graphcore Intelligence Processing Units (IPU). We evaluate the training performance on multiple molecular graph databases with varying degrees of graph counts, sizes and sparsity. We demonstrate that such a co-design approach can reduce the training time of such molecular property prediction models from days to less than two hours, opening new possibilities for AI-driven scientific discovery.

Vision-based perception tasks fulfill a paramount role in robotics, facilitating solutions to many challenging scenarios, such as acrobatics maneuvers of autonomous unmanned aerial vehicles (UAVs) and robot-assisted high precision surgery. Most control-oriented and egocentric perception problems are commonly solved by taking advantage of the robot state estimation as an auxiliary input, particularly when artificial intelligence comes into the picture. In this work, we propose to apply a similar approach for the first time - to the best of our knowledge - to allocentric perception tasks, where the target variables refer to an external subject. We prove how our general and intuitive methodology improves the regression performance of deep convolutional neural networks (CNNs) with ambiguous problems such as the allocentric 3D pose estimation. By analyzing three highly-different use cases, spanning from grasping with a robotic arm to following a human subject with a pocket-sized UAV, our results consistently improve the R2 metric up to +0.514 compared to their stateless baselines. Finally, we validate the in-field performance of a closed-loop autonomous pocket-sized UAV in the human pose estimation task. Our results show a significant reduction, i.e., 24% on average, on the mean absolute error of our stateful CNN.

Today, machine learning (ML) is widely used in industry to provide the core functionality of production systems. However, it is practically always used in production systems as part of a larger end-to-end software system that is made up of several other components in addition to the ML model. Due to production demand and time constraints, automated software engineering practices are highly applicable. The increased use of automated ML software engineering practices in industries such as manufacturing and utilities requires an automated Quality Assurance (QA) approach as an integral part of ML software. Here, QA helps reduce risk by offering an objective perspective on the software task. Although conventional software engineering has automated tools for QA data analysis for data-driven ML, the use of QA practices for ML in operation (MLOps) is lacking. This paper examines the QA challenges that arise in industrial MLOps and conceptualizes modular strategies to deal with data integrity and Data Quality (DQ). The paper is accompanied by real industrial use-cases from industrial partners. The paper also presents several challenges that may serve as a basis for future studies.

In this paper, we examine the concept of the predict and optimise problem with specific reference to the third Technical Challenge of the IEEE Computational Intelligence Society. In this competition, entrants were asked to forecast building energy use and solar generation at six buildings and six solar installations, and then use their forecast to optimize energy cost while scheduling classes and batteries over a month. We examine the possible effect of underforecasting and overforecasting and asymmetric errors on the optimisation cost. We explore the different nature of loss functions for the prediction and optimisation phase and propose to adjust the final forecasts for a better optimisation cost. We report that while there is a positive correlation between these two, more appropriate loss functions can be used to optimise the costs associated with final decisions.

The rapid progress of machine learning interatomic potentials over the past couple of years produced a number of new architectures. Particularly notable among these are the Atomic Cluster Expansion (ACE), which unified many of the earlier ideas around atom density-based descriptors, and Neural Equivariant Interatomic Potentials (NequIP), a message passing neural network with equivariant features that showed state of the art accuracy. In this work, we construct a mathematical framework that unifies these models: ACE is generalised so that it can be recast as one layer of a multi-layer architecture. From another point of view, the linearised version of NequIP is understood as a particular sparsification of a much larger polynomial model. Our framework also provides a practical tool for systematically probing different choices in the unified design space. We demonstrate this by an ablation study of NequIP via a set of experiments looking at in- and out-of-domain accuracy and smooth extrapolation very far from the training data, and shed some light on which design choices are critical for achieving high accuracy. Finally, we present BOTNet (Body-Ordered-Tensor-Network), a much-simplified version of NequIP, which has an interpretable architecture and maintains accuracy on benchmark datasets.

A parametric adaptive greedy Latent Space Dynamics Identification (gLaSDI) framework is developed for accurate, efficient, and certified data-driven physics-informed greedy auto-encoder simulators of high-dimensional nonlinear dynamical systems. In the proposed framework, an auto-encoder and dynamics identification models are trained interactively to discover intrinsic and simple latent-space dynamics. To effectively explore the parameter space for optimal model performance, an adaptive greedy sampling algorithm integrated with a physics-informed error indicator is introduced to search for optimal training samples on the fly, outperforming the conventional predefined uniform sampling. Further, an efficient k-nearest neighbor convex interpolation scheme is employed to exploit local latent-space dynamics for improved predictability. Numerical results demonstrate that the proposed method achieves 121 to 2,658x speed-up with 1 to 5% relative errors for radial advection and 2D Burgers dynamical problems.

During physical human robot collaboration, it is important to be able to implement a time-varying interactive behaviour while ensuring robust stability. Admittance control and passivity theory can be exploited for achieving these objectives. Nevertheless, when the admittance dynamics is time-varying, it can happen that, for ensuring a passive and stable behaviour, some spurious dissipative effects have to be introduced in the admittance dynamics. These effects are perceived by the user and degrade the collaborative performance. In this paper we exploit the task redundancy of the manipulator in order to harvest energy in the null space and to avoid spurious dynamics on the admittance. The proposed architecture is validated by simulations and by experiments onto a collaborative robot.

Computationally efficient and trustworthy machine learning algorithms are necessary for Digital Twin (DT) framework development. Generally speaking, DT-enabling technologies consist of five major components: (i) Machine learning (ML)-driven prediction algorithm, (ii) Temporal synchronization between physics and digital assets utilizing advanced sensors/instrumentation, (iii) uncertainty propagation, and (iv) DT operational framework. Unfortunately, there is still a significant gap in developing those components for nuclear plant operation. In order to address this gap, this study specifically focuses on the "ML-driven prediction algorithms" as a viable component for the nuclear reactor operation while assessing the reliability and efficacy of the proposed model. Therefore, as a DT prediction component, this study develops a multi-stage predictive model consisting of two feedforward Deep Learning using Neural Networks (DNNs) to determine the final steady-state power of a reactor transient for a nuclear reactor/plant. The goal of the multi-stage model architecture is to convert probabilistic classification to continuous output variables to improve reliability and ease of analysis. Four regression models are developed and tested with input from the first stage model to predict a single value representing the reactor power output. The combined model yields 96% classification accuracy for the first stage and 92% absolute prediction accuracy for the second stage. The development procedure is discussed so that the method can be applied generally to similar systems. An analysis of the role similar models would fill in DTs is performed.

In the research area of reinforcement learning (RL), frequently novel and promising methods are developed and introduced to the RL community. However, although many researchers are keen to apply their methods on real-world problems, implementing such methods in real industry environments often is a frustrating and tedious process. Generally, academic research groups have only limited access to real industrial data and applications. For this reason, new methods are usually developed, evaluated and compared by using artificial software benchmarks. On one hand, these benchmarks are designed to provide interpretable RL training scenarios and detailed insight into the learning process of the method on hand. On the other hand, they usually do not share much similarity with industrial real-world applications. For this reason we used our industry experience to design a benchmark which bridges the gap between freely available, documented, and motivated artificial benchmarks and properties of real industrial problems. The resulting industrial benchmark (IB) has been made publicly available to the RL community by publishing its Java and Python code, including an OpenAI Gym wrapper, on Github. In this paper we motivate and describe in detail the IB's dynamics and identify prototypic experimental settings that capture common situations in real-world industry control problems.

AI has already had a substantial positive influence in the battle against global warming. However, quantifying its significance and characterizing its effects remain open problems. This article provides an overview of AI-reliant initiatives and projects for understanding and combating climate change, highlights existing research documenting the potential positive impact of AI on climate change, and identifies a set of obstacles to be overcome to ensure such use of AI is both practical and ethically sound. Climate change will profoundly affect environments, societies, and economies. Many of its environmental repercussions, from prolonged droughts to more devastating storms, are already being seen.