Artificial intelligence is increasingly being integrated into socio-technical systems. Existing design principles for ethical, safe and trustworthy AI tend to be highly abstract and focus on AI systems in isolation. They have rarely considered the adverse effects on safety that may emerge from interactions between AI and other technical components. Organisational theories of safety take such emergent outcomes of interactions between entities in socio-technical systems into account. They offer guidance on how to identify structural vulnerabilities in socio-technical systems enhanced by AI, and how to organise the design and operation of such systems for safety. In this paper, which is the result of a collaboration between sociologists and computer scientists (AI consultants), we conduct an analysis that can support the process of designing AI-enhanced autonomous systems in order to avoid structural vulnerabilities. It builds on organisational theories of safety and derives five key descriptors from them, the examination of which can guide the design of AI-enhanced systems. We demonstrate the utility of the descriptors by applying them to proposals for AI-enhanced critical functions in advanced microgrids. We discuss these proposals from the research literature on microgrids and review their effects on structural vulnerabilities. We then explore the implications that go beyond the example of advanced microgrids and propose steps for reviewing and reflecting on structural vulnerabilities that AI controllers may introduce into socio-technical systems.

Deep learning (DL) is one of the key tools for analyzing images beyond the visible light spectrum, such as thermal data, for energy-related inspection and fault detection. However, publications using multispectral data focus on developing specialized models to handle quality issues without considering the imagery itself. This article investigates how feature engineering (FE), the process of adapting raw data to serve as DL training data, can impact performance when transferring prevalent model architectures to combined red, green, blue (RGB) thermal imagery. The popular U-Net is utilized for the common task of multiclass semantic segmentation in remote sensing. A comprehensive ablation study is performed on a novel, uncrewed aircraft system-based dataset from two German cities to detect thermal urban features. Common performance metrics, training, and energy consumption statistics are compared to find the most suitable combination of platform-specific and general enhancing FE while identifying the impact of resolution, channel count, RGB, and color information. The study reveals FE to significantly influence predictive performance, where the choice of ablation parameters are found to have a 7% -10% impact. Computational resource utilization depends on image size, following a logarithmic growth curve. Importantly, the study demonstrates that in-depth FE of thermal imagery can replace the need for additional RGB data.

Simulating high-resolution detector responses is a computationally intensive process that has long been challenging in Particle Physics. Despite the ability of generative models to streamline it, full ultra-high-granularity detector simulation still proves to be difficult as it contains correlated and fine-grained information. To overcome these limitations, we propose Intra-Event Aware Generative Adversarial Network (IEA-GAN). IEA-GAN presents a Transformer-based Relational Reasoning Module that approximates an event in detector simulation, generating contextualized high-resolution full detector responses with a proper relational inductive bias. IEA-GAN also introduces a Self-Supervised intra-event aware loss and Uniformity loss, significantly enhancing sample fidelity and diversity. We demonstrate IEA-GAN's application in generating sensor-dependent images for the ultra-high-granularity Pixel Vertex Detector (PXD), with more than 7.5 M information channels at the Belle II Experiment. Applications of this work span from Foundation Models for high-granularity detector simulation, such as at the HL-LHC (High Luminosity LHC), to simulation-based inference and fine-grained density estimation.

Simulating high-resolution detector responses is a storage-costly and computationally intensive process that has long been challenging in particle physics. Despite the ability of deep generative models to make this process more cost-efficient, ultra-high-resolution detector simulation still proves to be difficult as it contains correlated and fine-grained mutual information within an event. To overcome these limitations, we propose Intra-Event Aware GAN (IEA-GAN), a novel fusion of Self-Supervised Learning and Generative Adversarial Networks. IEA-GAN presents a Relational Reasoning Module that approximates the concept of an ”event” in detector simulation, allowing for the generation of correlated layer-dependent contextualized images for high-resolution detector responses with a proper relational inductive bias. IEA-GAN also introduces a new intra-event aware loss and a Uniformity loss, resulting in significant enhancements to image fidelity and diversity. We demonstrate IEA-GAN's application in generating sensor-dependent images for the high-granularity Pixel Vertex Detector (PXD), with more than 7.5M information channels and a non-trivial geometry, at the Belle II Experiment. Applications of this work include controllable simulation-based inference and event generation, high-granularity detector simulation such as at the HL-LHC (High Luminosity LHC), and fine-grained density estimation and sampling. To the best of our knowledge, IEA-GAN is the first algorithm for faithful ultra-high-resolution detector simulation with event-based reasoning.

Advanced cleaning strategies for parabolic trough collectors at concentrated solar power plants maximize the yield and minimize the costs for cleaning activities. However, they require information about the current soiling level of each collector. In this work, a novel, data-driven method for soiling estimation with machine learning for parabolic trough collectors is developed using gloss values as a surrogate for soiling values. Operational data and meteorological data from the solar field Andasol-3 with changing time horizons are used together with various Machine Learning techniques to estimate the soiling of every collector in the field. The best results were achieved with a Decision Tree model, with a coefficient of determination of R2=0.77 from the maximum value of 1 and a mean squared error of MSE=6.14 for the determination of specific soiling values. A second metric to evaluate the quality of soiling predictions from the models classifies whether soiling is above or below a cleaning threshold was also investigated. Model results are compared to soiling measurements that indicate the need for cleanings. Cleaning recommendations are derived and compared with the current fixed-time cleaning schedule of Andasol-3. All models show an improvement over the cleaning schedule currently in use. The use of a Decision Tree model increases the detected necessary cleanings by 12.2 %, while the number of unnecessary cleanings are reduced by 14.3 %. This has the potential to reduce operational costs and increase the solar field yield. The dataset used in this work is made publicly available https://doi.org/10.5281/zenodo.7061913, along with the code to reproduce all results, which can be found at https://doi.org/10.5281/zenodo.7554806.

Thermal bridges are weak points of building envelopes that can lead to energy losses, collection of moisture, and formation of mould in the building fabric. To detect thermal bridges of large building stocks, drones with thermographic cameras can be used. As the manual analysis of comprehensive image datasets is very time-consuming, we investigate deep learning approaches for its automation. For this, we focus on thermal bridges on building rooftops recorded in panorama drone images from our updated dataset of Thermal Bridges on Building Rooftops (TBBRv2), containing 926 images with 6,927 annotations. The images include RGB, thermal, and height information. We compare state-of-the-art models with and without pretraining from five different neural network architectures: MaskRCNN R50, Swin-T transformer, TridentNet, FSAF, and a MaskRCNN R18 baseline. We find promising results, especially for pretrained models, scoring an Average Recall above 50% for detecting large thermal bridges with a pretrained Swin-T Transformer model.

Thermal Bridges on Building Rooftops (TBBR) is a multi-channel remote sensing dataset. It was recorded during six separate UAV fly-overs of the city center of Karlsruhe, Germany, and comprises a total of 926 high-resolution images with 6927 manually-provided thermal bridge annotations. Each image provides five channels: three color, one thermographic, and one computationally derived height map channel. The data is pre-split into training and test data subsets suitable for object detection and instance segmentation tasks. All data is organized and structured to comply with FAIR principles, i.e. being findable, accessible, interoperable, and reusable. It is publicly available and can be downloaded from the Zenodo data repository. This work provides a comprehensive data descriptor for the TBBR dataset to facilitate broad community uptake.

In this work, we present a neural approach to reconstructing rooted tree graphs describing hierarchical interactions, using a novel representation we term the lowest common ancestor generations (LCAG) matrix. This compact formulation is equivalent to the adjacency matrix, but enables learning a tree’s structure from its leaves alone without the prior assumptions required if using the adjacency matrix directly. Employing the LCAG therefore enables the first end-to-end trainable solution which learns the hierarchical structure of varying tree sizes directly, using only the terminal tree leaves to do so. In the case of high-energy particle physics, a particle decay forms a hierarchical tree structure of which only the final products can be observed experimentally, and the large combinatorial space of possible trees makes an analytic solution intractable. We demonstrate the use of the LCAG as a target in the task of predicting simulated particle physics decay structures using both a Transformer encoder and a neural relational inference encoder graph neural network. With this approach, we are able to correctly predict the LCAG purely from leaf features for a maximum tree-depth of 8 in 92.5 % of cases for trees up to 6 leaves (including) and 59.7 % for trees up to 10 in our simulated dataset.

With increasing data and model complexities, the time required to train neural networks has become prohibitively large. To address the exponential rise in training time, users are turning to data parallel neural networks (DPNN) and large-scale distributed resources on computer clusters. Current DPNN approaches implement the network parameter updates by synchronizing and averaging gradients across all processes with blocking communication operations after each forward-backward pass. This synchronization is the central algorithmic bottleneck. We introduce the Distributed Asynchronous and Selective Optimization (DASO) method, which leverages multi-GPU compute node architectures to accelerate network training while maintaining accuracy. DASO uses a hierarchical and asynchronous communication scheme comprised of node-local and global networks while adjusting the global synchronization rate during the learning process. We show that DASO yields a reduction in training time of up to 34% on classical and state-of-the-art networks, as compared to current optimized data parallel training methods.

. Thermal bridges are weak areas of building envelopes that conduct more heat to the outside than surrounding envelope areas. They lead to increased energy consumption and the formation of mold. With a neural network approach, we demonstrate a method of automatically detecting thermal bridges on building rooftops from panorama drone images of whole city districts. To train the neural network, we created a dataset including 917 images and 6895 annotations. The images in the dataset contain thermal information for detecting thermal bridges and a height map for rooftop recognition in addition to regular RGB information. Due to the small dataset, our approach currently only has an average recall of 9.4% @IoU:0.5-0.95 (14.4% for large objects). Nevertheless, our approach reliably detects structures only on rooftops and not on other parts of buildings, without any additional segmentation effort of building parts.

Statistical testing is widespread and critical for a variety of scientific disciplines. The advent of machine learning and the increase of computing power has increased the interest in the analysis and statistical testing of multidimensional data. We extend the powerful Kolmogorov-Smirnov two sample test to a high dimensional form in a similar manner to Fasano (Fasano, 1987). We call our result the d-dimensional Kolmogorov-Smirnov test (ddKS) and provide three novel contributions therewith: we develop an analytical equation for the significance of a given ddKS score, we provide an algorithm for computation of ddKS on modern computing hardware that is of constant time complexity for small sample sizes and dimensions, and we provide two approximate calculations of ddKS: one that reduces the time complexity to linear at larger sample sizes, and another that reduces the time complexity to linear with increasing dimension. We perform power analysis of ddKS and its approximations on a corpus of datasets and compare to other common high dimensional two sample tests and distances: Hotelling's T^2 test and Kullback-Leibler divergence. Our ddKS test performs well for all datasets, dimensions, and sizes tested, whereas the other tests and distances fail to reject the null hypothesis on at least one dataset. We therefore conclude that ddKS is a powerful multidimensional two sample test for general use, and can be calculated in a fast and efficient manner using our parallel or approximate methods. Open source implementations of all methods described in this work are located at https://github.com/pnnl/ddks.

The Belle II experiment, beginning data taking with the full detector in early 2019, is expected to produce a volume of data ffty times that of its predecessor. This dramatic increase in data comes the opportunity for studies of rare previously inaccessible processes. The investigation of such rare processes in a high data-volume environment requires a correspondingly high volume of Monte Carlo simulations to prepare analyses and gain a deep understanding of the contributing physics processes to each individual study. This presents a signifcant challenge in terms of computing resource requirements and calls for more intelligent methods of simulation, in particular background processes with very high rejection rates. This work presents a method of predicting in the early stages of the simulation process the likelihood of relevancy of an individual event to the target study using convolutional neural networks. The results show a robust training that is integrated natively into the existing Belle II analysis software framework.

We present a d-dimensional test statistic inspired directly by the Kolmogorov– Smirnov (KS) test statistic and Press’ extension of the KS test to two dimensions. We call this the ddKS statistic. To preclude the high computational cost associated with working in higher dimensions, we present an implementation using tensor primitives. This allows parallel computation on CPU or GPU. We explore the behavior of the test statistic in comparing two three-dimensional samples, and use a standard statistical method the permutation method to explore its significance. We show that, while the Kullback–Leibler divergence is a good choice for general distribution comparison, ddKS has properties that make it more desirable for surrogate model training and validation than Kullback–Leibler divergence. Diverse fields of research, especially the physical sciences and studies including surrogate modeling and normalizing flows, require the comparison of high-dimensional distributions. In many cases, analysts of these distributions revert to using one dimensional comparison tools, such as the Kolmogorov–Smirnov test, simple formulation of the Earth Mover’s Distance, or even the mean integrated squared error between histograms. These one-dimensional comparisons suffer from the inability to identify differences in the covariances encoded by each distribution. Therefore, efficient (both statistically and computationally) comparisons between high-dimensional distributions are essential to advancing the use of machine learning and statistical analysis in the physical sciences. We present one such technique. One of the most used, and arguably most powerful, two-sample tests is that proposed by Kolmogorov and tabulated by Smirnov, the so-called Kolmogorov–Smirnov (KS) test (1; 2). This test is widely used to compare two samples to determine whether they come from the same one-dimensional distribution. It does so by creating a test statistic, which is the maximum difference between the cumulative density function of two samples. The significance can then be easily calculated from that test statistic, regardless of whether either sample came from a well-defined distribution or not. This test has several important properties for machine learning, including its fast computation and ability to test non-parametric samples. However, it comes with one major drawback: that it is only applicable in one dimension, whereas many problems in data science cannot be compressed to one dimension without loss of information. Several authors, mostly notably Press et al. (3), have defined similar tests in higher dimensions. The difficulty in extension to many dimensions is the ambiguity in the cumulative density function for a many dimensional distribution. Press et al. define a test statistic in two dimensions using the class membership in each of the four quadrants surrounding each test point. The maximum of the differences between membership vectors becomes a test statistic very similar to that in the KS test, however the same properties for calculating the significance are not retained. Press’ test is of polynomial time complexity, and perhaps due to this, has not seen wide acceptance (or possibly awareness) in the data science community. Third Workshop on Machine Learning and the Physical Sciences (NeurIPS 2020), Vancouver, Canada. We present a d-dimensional KS test, inspired directly by Press’ methodology of extension, but we present two novel changes. Firstly, we present a tensor-based calculation of the statistic, which reduces the computational complexity for small sample sizes. Secondly, we use the permutation method to define significance for our examples, and illustrate how it could be used for practicing data scientists with our test statistic. 1 The d-dimensional Kolmogorov–Smirnov test statistic We begin with two samples (a predicted, Xp, and true, Xt, sample) of N points, each point having d dimensions. We seek to compare the cumulative density functions (CDF) between these samples. The construction of a CDF is ambiguous in more than one dimension, however an often used surrogate is the membership in hyperspace regions partitioned at a given test point. In one dimension, this is equivalent to choosing a test point, and measuring two numbers: counting membership greater than and less than the given test point. This concept generalizes to many dimensions by using each test point as an origin to delineate regions in hyperspace and forming a 2 membership vector whose components correspond to the number of points in each region. The ddKS distance using points from Xp to partition between the two samples is designated Dp. Let xi ∈ Xp and V p j (xi), V t j (xi) be the jth component of the membership vectors for the predicted and true samples generated by the partition of space due to the point xi. The ddKS divergence between Xp and Xt using Xp is the largest element of the difference of the two membership vectors: Dp = max i,j |V t j (xi)− V p j (xi)| (1) Where the subscript p indicates use of the predicted dataset as the partitioning points; a subscript t indicates use of the true dataset as the partitioning points. The ddKS test statistic, D, is then the average of Dp and Dt. The most straightforward implementation of ddKS can be constructed using loop-based logic. For each point in each sample, the membership of the hyperspace regions using that point as a partition is counted. As described above, we evaluate the region membership of both, using predicted and true samples as partitioning points, and average the maximum differences of their membership vectors. The naive loop-based implementation has high computational cost, on the order of O ( 2N ) for all N . Because of this, we have developed a tensor-based implementation which utilizes pytorch’s (4) implicit parallelization on CPU or GPU. The computational complexity of the tensor form of ddKS is O ( 2 ) for small N , with memory complexity O ( N ) . We implemented this tensor-based method with pytorch, enabling use of modern GPU hardware.

The large volume of data expected to be produced by the Belle II experiment presents the opportunity for studies of rare, previously inaccessible processes. Investigating such rare processes in a high data volume environment necessitates a correspondingly high volume of Monte Carlo simulations to prepare analyses and gain a deep understanding of the contributing physics processes to each individual study. This resulting challenge, in terms of computing resource requirements, calls for more intelligent methods of simulation, in particular for processes with very high background rejection rates. This work presents a method of predicting in the early stages of the simulation process the likelihood of relevancy of an individual event to the target study using graph neural networks. The results show a robust training that is integrated natively into the existing Belle II analysis software framework.

The Full Event Interpretation is presented: a new exclusive tagging algorithm used by the high-energy physics experiment Belle II. The experimental setup of Belle II allows the precise measurement of otherwise inaccessible $B$ meson decay-modes. The Full Event Interpretation algorithm enables many of these measurements. The algorithm relies on machine learning to automatically identify plausible $B$ meson decay chains based on the data recorded by the detector. Compared to similar algorithms employed by previous experiments, the Full Event Interpretation provides a greater efficiency, yielding a larger effective sample size usable in the measurement.