Vision Language Models (VLMs) have shown remarkable capabilities in multimodal under-standing, yet their susceptibility to adversarial perturbations poses a significant threat to their reliability in real-world applications. Despite often being imperceptible to humans, these perturbations can drastically alter model outputs, leading to erroneous interpretations and decisions. This paper introduces DiffCAP, a novel diffusion-based purification strategy that can effectively neutralize adversarial corruptions in VLMs. We theoretically establish a provable recovery region in the forward diffusion process and meanwhile quantify the convergence rate of semantic variation with respect to VLMs. These findings manifest that adversarial effects monotonically fade as diffusion unfolds. Guided by this principle, DiffCAP leverages noise injection with a similarity threshold of VLM embeddings as an adaptive criterion, before reverse diffusion restores a clean and reliable representation for VLM inference. Through extensive experiments across six datasets with three VLMs under varying attack strengths in three task scenarios, we show that DiffCAP outperforms existing defense techniques by a substantial margin. Notably, DiffCAP significantly reduces both hyperparameter tuning complexity and the required diffusion time, thereby accelerating the denoising process. Equipped with theorems and empirical support, DiffCAP provides a robust and practical solution for securely deploying VLMs in adversarial environments. The source code is available at https://github.com/JasonFu1998/DiffCAP

Anomaly detection is an essential component for ensuring the safety and reliability of critical systems. Currently, most machine learning-based anomaly detection approaches rely purely on correlations among sensor signals, rather than causal relations, making them susceptible to spurious associations. This limitation can lead to poor generalization and unreliable anomaly detection in practical scenarios. To address this, we propose an anomaly detection approach which leverages score-based causal discovery, NTS-DAGMA. This causal discovery method advances over prior work by combining the network architecture from NTS-NOTEARS with the acyclicity constraint from DAGMA. Like other prediction-based anomaly detection methods, it can predict future states; however, it does this by learning and utilizing causal relations in the time series data. Through comprehensive experiments, we demonstrate that: (i) on causal discovery tasks NTS-NOTEARS and NTS-DAGMA achieve similar performances; (ii) on anomaly detection tasks NTS-NOTEARS and NTS-DAGMA also perform similar to each other, and have comparable performance to state-of-the-art ML approaches; and (iii) most importantly, our results show that NTS-DAGMA provides causally meaningful models: detects anomalies which propagate through child nodes in agreement with the inferred causal graph. Incorporating causal structure into the model enables improved interpretability and aligns anomaly detection with the physical dynamics of the system

In the ever-evolving adversarial machine learning landscape, developing effective defenses against patch attacks has become a critical challenge, necessitating reliable solutions to safeguard real-world AI systems. Although diffusion models have shown remarkable capacity in image synthesis and have been recently utilized to counter lp-norm bounded attacks, their potential in mitigating localized patch attacks remains largely underexplored. In this work, we propose DiffPAD, a novel framework that harnesses the power of diffusion models for adversarial patch decontamination. DiffPAD first performs super-resolution restoration on downsampled input images, then adopts binarization, dynamic thresholding scheme and sliding window for effective localization of adversarial patches. Such a design is inspired by the theoretically derived correlation between patch size and diffusion restoration error that is generalized across diverse patch attack scenarios. Finally, DiffPAD applies inpainting techniques to the original input images with the estimated patch region being masked. By integrating closed-form solutions for super-resolution restoration and image inpainting into the conditional reverse sampling process of a pre-trained diffusion model, DiffPAD obviates the need for text guidance or fine-tuning. Through comprehensive experiments, we demonstrate that DiffPAD not only achieves state-of-the-art adversarial robustness against patch attacks but also excels in recovering naturalistic images without patch remnants. The source code is available at https://github.com/JasonFu1998/DiffPAD. 

Self-driving labs (SDLs) employing automation and machine learning (ML) offer great promise for accelerating materials discovery and optimisation. However, in thin film science, SDLs are mainly restricted to solution-based methods which are easier to automate, restricting access to the broader chemical space of inorganic materials. This work advances an SDL based on magnetron co-sputtering, addressing a key challenge: rapidly generating accurate composition maps of multi-element, compositionally graded thin films. Traditional ex-situ methods are slow and error-prone; instead, we present a fast, calibration-free, in-situ ML approach to predict the deposition rate using quartz-crystal microbalance (QCM) sensors. For each sputtering source, deposition rates are sequentially learned as a function of pressure and power via active learning with Gaussian processes (GPs). The final GPs are combined with a geometric flux model to interpolate deposition rates across the sample. Among several acquisition functions with random query as the baseline, the Bayesian active learning MacKay (BALM) approach yielded the best performance, requiring as few as 10 experiments per source. The model predictions for co-sputtering composition distributions were validated against external composition measurements. This framework significantly increases throughput in combinatorial sputtering studies and highlights the potential of ML-guided SDLs to surpass traditional Edisonian methods

We describe the AI-ARC (Artificial Intelligence-based Virtual Control Room for the Arctic) system, which aims to enhance maritime domain awareness and surveillance. The system is micro-service based and fuses data from various sources, utilizing AI-driven micro-services and an advanced visualization platform to increase the situation awareness of maritime surveillance operators. The results of the Baltic sea demonstration, aiding in the detection of illegal activities, environmental protection, are presented. The system was evaluated using historical data from real criminal incidents. The resultsshow that the AI-ARC approach could help increase the situation awareness of law enforcement operators. 

Linear potential flow (LPF) models remain the tools-of-the trade in marine and ocean engineering despite their well-known assumptions of small amplitude waves and motions. As of now, nonlinear simulation tools are still too computationally demanding to be used in the entire design loop, especially when it comes to the evaluation of numerous irregular sea states. In this paper we aim to enhance the performance of the LPF models by introducing a hybrid LPF-ML (machine learning) approach, based on identification of nonlinear force corrections. The corrections are defined as the difference in hydrodynamic force (vis- cous and pressure-based) between high-fidelity CFD and LPF models. Using prescribed chirp motions with different amplitudes, we train a long short-term memory (LSTM) network to predict the corrections. The LSTM network is then linked to the MoodyMarine LPF model to provide the nonlinear correction force at every time-step, based on the dynamic state of the body and the corresponding forces from the LPF model. The method is illustrated for the case of a heaving sphere in decay, regular and irregular waves – including passive control. The hybrid LPF-model is shown to give significant improvements compared to the baseline LPF model, even though the training is quite generic.

Numerical models used in the design of floating bodies routinely rely on linear hydrodynamics. Extensions for hydrodynamic nonlinearities can be approximated using e.g. Morison type drag and nonlinear Froude-Krylov forces. This paper aims to improve the approximation of nonlinear forces acting on floating bodies by using machine learning (ML). Many ML models are general function approximators and therefore suitable for representing such nonlinear correction terms. A hierarchical modelling approach is used to build mappings between higher-fidelity simulations and the linear method. The ML corrections are built up for FNPF, Euler and RANS simulations. Results for decay tests of a sphere in model scale using recurrent neural networks (RNN) are presented. The RNN algorithm is shown to satisfactory predict the correction terms if the most nonlinear case is used as training data. No difference in the performance of the RNN model is seen for the different hydrodynamic models.

We present a hybrid linear potential flow - machine learning (LPF-ML) model for simulating weakly nonlinear wave-body interaction problems. In this paper we focus on using hierarchical modelling for generating training data to be used with recurrent neural networks (RNNs) in order to derive nonlinear correction forces. Three different approaches are investigated: (i) a baseline method where data from a Reynolds averaged Navier Stokes (RANS) model is directly linked to data from a LPF model to generate nonlinear corrections; (ii) an approach in which we start from high-fidelity RANS simulations and build the nonlinear corrections by stepping down in the fidelity hierarchy; and (iii) a method starting from low-fidelity, successively moving up the fidelity staircase. The three approaches are evaluated for the simple test case of a heaving sphere. The results show that the baseline model performs best, as expected for this simple test case. Stepping up in the fidelity hierarchy very easily introduce errors that propagate through the hierarchical modelling via the correction forces. The baseline method was found to accurately predict the motion of the heaving sphere. The hierarchical approaches struggled with the task, with the approach that steps down in fidelity performing somewhat better of the two.

Maintenance and services of products as well as processes are pivotal for achieving high availability and avoiding catastrophic and costly failures. At the same time, maintenance is routinely performed more frequently than necessary, replacing possibly functional components, which has negative economic impact on the maintenance. New processes and products need to fulfil increased environmental demands, while customers put increasing demands on customization and coordination. Hence, improved maintenance processes possess very high potentials, economically as well as environmentally. The shifting demands on product development and production processes have led to the emergency of new digital solutions as well as new business models, such as integrated product-service offerings. Still, the general maintenance problem of how to perform the right service at the right time, taking available information and given limitations is valid. The project Future Industrial Services Management (FUSE) project was a step in a long-term effort for catalysing the evolution of maintenance and production in the current digital era. In this paper, several aspects of the general maintenance problem are discussed from a data driven perspective, spanning from technology solutions and organizational requirements to new business opportunities and how to create optimal maintenance plans. One of the main results of the project, in the form of a simulation tool for strategy selection, is also described.

Background: Understanding temporal patterns of organ dysfunction (OD) may aid early recognition of complications after trauma and assist timing and modality of treatment strategies. Our aim was to analyse and characterise temporal patterns of OD in intensive care unit-admitted trauma patients. Methods: We used group-based trajectory modelling to identify temporal trajectories of OD after trauma. Modelling was based on the joint development of all six subdomains comprising the sequential organ failure assessment score measured daily during the first two weeks post trauma. Further, the time for trajectories to stabilise and transition to final group assignments were evaluated. Results: Six-hundred and sixty patients were included in the final model. Median age was 40 years, and median ISS was 26 (IQR 17–38). We identified five distinct trajectories of OD. Group 1, mild OD (n = 300), median ISS of 20 (IQR 14–27), had an early resolution of OD and a low mortality. Group 2, moderate OD (n = 135), and group 3, severe OD (n = 87), were fairly similar in admission characteristics and initial OD but differed in subsequent OD trajectories, the latter experiencing an extended course and higher mortality. In group 3, 56% of the patients developed sepsis as compared with 19% in group 2. Group 4, extreme OD (n = 40), received most blood transfusions, had the highest proportion of shock at admission and a median ISS of 41 (IQR 29–50). They experienced significant and sustained OD affecting all organ systems and a 28-day mortality of 30%. Group 5, traumatic brain injury with OD (n = 98), had the highest mortality of 35% and the shortest time to death for non-survivors, median 3.5 (IQR 2.4–4.8) days. Groups 1 and 5 reached their final group assignment early, > 80% of the patients within 48 h. In contrast, groups 2 and 3 had a prolonged time to final group assignment. Conclusions: We identified five distinct trajectories of OD after severe trauma during the first two weeks post-trauma. Our findings underline the heterogeneous course after trauma and describe some potentially important clinical insights that are suggested by the groupings and temporal trajectories. © 2021, The Author(s).