To raise awareness of the environmental impact of deep learning (DL), numerous studies have estimated the energy consumption of DL systems. However, energy estimates during DL training often rely on unverified assumptions. This work addresses that gap by investigating how model architecture and training environment affect energy consumption. We train a variety of computer vision models and collect energy consumption and accuracy metrics to analyze their trade-offs across configurations. Our results show that selecting the right model-training environment combination can reduce training energy consumption by up to 80.68% with less than 2% loss in F1 score. We find a significant interaction effect between model and training environment: energy efficiency improves when GPU computational power scales with model complexity. Moreover, we demonstrate that common estimation practices, such as using FLOPs or GPU TDP, fail to capture these dynamics and can lead to substantial errors. To address these shortcomings, we propose the Stable Training Epoch Projection (STEP) and the Pre-training Regression-based Estimation (PRE) methods. Our evaluation demonstrates that STEP and PRE achieve reductions in Root Mean Squared Error (RMSE) up to 97% and 84%, respectively, when compared to existing estimation tools.
Ammonia is a promising carbon–neutral fuel for solid oxide fuel cells (SOFCs), but its direct use can cause degradation due to nitridation. Pre-cracking ammonia reduces this issue but increases system complexity and energy consumption. This study examines both direct and pre-cracked ammonia as fuels, focusing on cracking kinetics across various temperatures and flow rates (3.4 − 50 mL min⁻1 cm⁻2). A 3D multiphysics model is developed, incorporating mass transport, electrochemical reactions, and ammonia decomposition, and shows strong agreement with experimental data between 700 °C and 850 °C. Results indicate nearly complete ammonia decomposition (>99 %) above 700 °C, except at high flow rates, with additional cracking occurring outside the cell due to reactions with metallic components. At lower temperatures and higher flow rates, incomplete decomposition reduces performance and may lead to Ni nitridation. Although direct ammonia yields lower performance, efficiency improves with higher temperature and fuel utilization, highlighting the importance of optimizing operating conditions.
_Latent-space_ monitoring techniques have shown promise as defenses against LLM attacks. These defenses act as scanners to detect harmful activations before they lead to undesirable actions. This prompts the question: can models execute harmful behavior _via inconspicuous latent states_? Here, we study such _obfuscated activations_. Our results are nuanced. We show that state-of-the-art latent-space defenses---such as activation probes and latent OOD detection---are vulnerable to obfuscated activations. For example, against probes trained to classify harmfulness, our obfuscation attacks can reduce monitor recall from 100% down to 0% while still achieving a 90% jailbreaking success rate. However, we also find that certain probe architectures are more robust than others, and we discover the existence of an _obfuscation tax_: on a complex task (writing SQL code), evading monitors reduces model performance. Together, our results demonstrate white-box monitors are not robust to adversarial attack, while also providing concrete suggestions to alleviate, but not completely fix, this weakness.
The tens of millions of spectra being captured by the Dark Energy Spectroscopic Instrument (DESI) provide tremendous discovery potential. In this work we show how Machine Learning, in particular Variational Autoencoders (VAE), can detect anomalies in a sample of approximately 200,000 DESI spectra comprising galaxies, quasars and stars. We demonstrate that the VAE can compress the dimensionality of a spectrum by a factor of 100, while still retaining enough information to accurately reconstruct spectral features. We then detect anomalous spectra as those with high reconstruction error and those which are isolated in the VAE latent representation. The anomalies identified fall into two categories: spectra with artefacts and spectra with unique physical features. Awareness of the former can help to improve the DESI spectroscopic pipeline; whilst the latter can lead to the identification of new and unusual objects. To further curate the list of outliers, we use the Astronomaly package which employs Active Learning to provide personalised outlier recommendations for visual inspection. In this work we also explore the VAE latent space, finding that different object classes and subclasses are separated despite being unlabelled. We demonstrate the interpretability of this latent space by identifying tracks within it that correspond to various spectral characteristics. For example, we find tracks that correspond to increasing star formation and increase in broad emission lines along the Balmer series. In upcoming work we hope to apply the methods presented here to search for both systematics and astrophysically interesting objects in much larger datasets of DESI spectra.
Palletizing in intralogistics and manufacturing remains particularly challenging when dealing with irregularly shaped products, large variant sets, and frequently changing customer orders. This paper presents an AI-enabled digital twin framework for reconfigurable robotic palletizing of irregularly shaped products. At its core, the framework uses a digital twin that maintains a unified representation of the robotic cell, its resources, and the handled products, and exposes consistent information to all software modules, enabling their interoperation. The palletizing pipeline integrates key enabling technologies to answer how products should be handled, where they are located in the workspace, and how they should be stacked at the output station. First, geometry-based preprocessing and an analytical grasping procedure extract key attributes and grasp points for each product type and store them in a knowledge base. Second, a perception module combines deep learning–based instance segmentation with template matching to identify and localize aluminium profiles that arrive in mixed, colour-oriented stacks without explicit identifiers. Third, a pallet loading optimizer employs a deep reinforcement learning–assisted selector, a constraint satisfaction stage, and a multi-criteria placement evaluation to generate near-optimal stacking configurations that satisfy stability and handling constraints. The digital twin mediates the exchange of this information across the participating enabling technologies, serving as a single source of truth and dispatching the solution execution components responsible for materializing the generated palletizing plan. The framework is deployed and validated in an industrial aluminium extrusion packaging use case using a dual-arm robotic palletizing cell, illustrating a viable path toward reconfigurable automation of irregular product palletizing in intralogistics and manufacturing environments.
