The Alan Turing Institute

Policy makers at local, national, and international levels are increasingly being required to make decisions that mitigate the effects of climate change on society and the economy. Earth Observations (EO) are already an important source of data to support such decisions, but this data represents only one aspect of the broader socio-technical systems that decision makers seek to influence. Policy effectiveness depends not only on environmental conditions, but also on household behaviour, technology adoption decisions, economic constraints, and feedbacks across scales. Capturing these dynamics requires modelling approaches that explicitly represent human decision-making alongside EO-derived inputs. This paper will present the results from the development of a scenario-planning digital twin (SPDT) designed to support decision-making processes related to local energy policies. The new SPDT will demonstrate how EO datasets can be integrated with multilevel agent-based models (MABMs) to enable specific scenarios to be used to support policy decisions. The use case in this work enables policy makers to (i) model residential heating demand, (ii) test policy levers that might best encourage the uptake of low-carbon heating, and (iii) assess the implications for energy use and fuel poverty. Specifically, the MABM simulates hourly residential energy demand for space heating at the household level, accounting for building characteristics, policy levers, occupancy patterns, retail energy prices, and external ambient temperature. The MABM supports baseline demand estimation at fine spatial granularity (individual households), the assessment of new technologies (such as retrofit measures or heating controls/meters) and energy price variations, and counterfactual analysis (e.g., setpoint shifts, tariff changes, warm/cold snaps). Earth observation data is used to inform medium to long-term climate trends for the use case region. The project results presented in this paper have been developed to serve the city of Newcastle upon Tyne in the UK. We discuss results for Newcastle developed so far, and possible new future scenarios that could be developed using this type of methodology.

Abstract The Manifold Hypothesis is a widely accepted tenet of Machine Learning which asserts that nominally high-dimensional data are in fact concentrated near a low-dimensional manifold, embedded in high-dimensional space. This phenomenon is observed empirically in many real world situations, has led to development of a wide range of statistical methods in the last few decades, and has been suggested as a key factor in the success of modern AI technologies. We show that rich and sometimes intricate manifold structure in data can emerge from a generic and remarkably simple statistical model — the Latent Metric Space model — via elementary concepts such as latent variables, correlation and stationarity. This establishes a general statistical explanation for why the Manifold Hypothesis seems to hold in so many situations. Informed by the Latent Metric Space model we derive procedures to discover and interpret the geometry of high-dimensional data, and explore hypotheses about the data generating mechanism. These procedures operate under minimal assumptions and make use of well-known dimension reduction methods and graph-analytic algorithms.

Supply Chain Management relies on human consensus in decision-making to avoid emergent problems like the bullwhip effect. Some routine consensus processes, especially those that are time-intensive, can be automated. Previously proposed supply chain automation solutions for consensus-seeking and coordination faced computational challenges, resulting in high entry barriers. Recent advances in Generative AI, particularly Large Language Model agents (LLM agents), could overcome these barriers. This paper explores how LLM agents can automate consensus-seeking in supply chains. We introduce a series of novel, supply chain-specific consensus-seeking frameworks and validate the effectiveness of our approach through a case study in inventory management, where agents that represent companies in a supply chain are able to balance selfish goals with systemic outcomes through conversation. Our results show that introducing LLM-based consensus-seeking frameworks reduces bullwhip effects. When equipped with appropriate tools, LLM agents can minimise bullwhip better than restocking policies and centralised demand approaches. Additionally, when LLM agents are handled within a negotiation framework, their behaviour converges to best practices in the supply chain literature on how to lessen the bullwhip effect. To provide a foundation for further advancements in LLM-based autonomous supply chain solutions, we open-source our code.

As a direct consequence of liquid kerosene injection, aeroengine combustors may be categorized as non-premixed combustion systems, characterized by a swirl-stabilized and highly complex flow field. In addition to the flow of air through the fuel injector, there are a large number of other features through which oxidizer can enter the heat release region. These can have an impact on local fuel-air mixing, inducing strong spatial and temporal variations in stoichiometry, thereby affecting emissions and combustion system performance. This paper discusses a novel statistical methodology, based on Principal Component Analysis (PCA) and K-means clustering, that aims to improve understanding of fuel-air mixing in realistic aeroengine combustors. The method is applied in a postprocessing step to data sampled from a Large Eddy Simulation (LES), where every chamber inflow has been tagged with a unique passive scalar, which allows it to be traced across space and time. PCA is used to construct a low-dimensional, visually interpretable representation of a spatially localized fuel-air mixing process, while K-means clustering is employed to produce an unsupervised discretization of the flow field into regions of similar fuel-air mixing characteristics. The proposed methodology is computationally inexpensive, and the easily interpretable outputs can help the combustion engineer make better informed decisions about combustor design.

Introduction Digital identity infrastructures used in electronic passports, national eID schemes, and federated authentication systems rely predominantly on centralised registries and classical public key cryptography. These architectures enable large-scale identity correlation, mass data aggregation, and single points of compromise, while remaining vulnerable to quantum attacks against RSA and elliptic-curve cryptography. There is no deployed identity framework that simultaneously provides post-quantum security, cryptographic privacy guarantees, and decentralised trust. Methods This study proposes a quantum-proof digital passport architecture combining lattice-based post-quantum cryptography, decentralised blockchain identifiers, and transformer-based decentralised artificial intelligence. The framework employs NIST-aligned post-quantum key encapsulation and digital signatures, zero-knowledge proofs for selective disclosure of identity attributes, and homomorphic encryption for encrypted identity verification. Blockchain oracles and decentralised identifiers enforce credential integrity and auditability without reliance on central identity providers. Transformer attention mechanisms support adaptive identity validation while preventing persistent identity profiling. Results Architectural analysis shows that the proposed system prevents quantum-enabled credential forgery, retrospective decryption, and cross-service identity linkability. Zero-knowledge verification removes plaintext exposure of personal data, and decentralised credential control eliminates central compromise vectors. The design remains interoperable with existing passport and eID infrastructures. Discussion The results demonstrate that secure post-quantum digital identity requires the combined application of quantum-resistant cryptography, decentralised governance, and cryptographic privacy enforcement.

ObjectivesLittle is known about how young people use social media during periods of self-harm. This study aimed to explore how they express themselves online through images posted on social media before and after self-harm and how this expression may change across these periods, employing visual content and thematic analyses.DesignA prospective cohort study, with qualitative analysis conducted using a recurrent cross-sectional approach and codebook methodology, accounting for chronological changes across time points before, during and after episodes of self-harm.SettingParticipants were recruited from a mental health NHS Trust in the UK.ParticipantsImage data during episodes of self-harm was available for 20 participants. The majority of whom were aged 18 years or older (n=15), female (n=14) and met criteria for moderate or severe anxiety and depression (n=18). The sample reflected diverse ethnic backgrounds, with six participants identifying as Asian or Mixed/Multiple ethnic backgrounds.ResultsNone of the images investigated had direct visual presentations of self-harm. A few images referenced self-harm through the medium of text, and this was largely to normalise and promote help-seeking. Several themes were identified, including participation in activities that support well-being, love and relationships, connecting through humour, expressions of distress, and promoting mental health awareness and support. Subtle temporal changes were also observed.ConclusionsFindings suggest that young people may temporarily withdraw from social media on the day of a self-harm event and rarely post graphic self-harm images around that time. This may reflect concerns about being stigmatised, but also improved platform moderation. Instead, platforms may serve as spaces for expressing self-care behaviours and connecting with others about both positive and challenging emotions, and across a range of topics including mental health.Trial registration numberClinicalTrials.gov: NCT04601220.

Deep learning based diagnostic AI systems based on medical images are starting to provide similar performance as human experts. However, these data-hungry complex systems are inherently black boxes and therefore slow to be adopted for high-risk applications like healthcare. This problem of lack of transparency is exacerbated in the case of recent large foundation models, which are trained in a self-supervised manner on millions of data points to provide robust generalisation across a range of downstream tasks. The embeddings generated from them happen through a process that is not interpretable, and hence not easily trustable for clinical applications. To address this timely issue, we deploy conformal analysis to quantify the predictive uncertainty of a vision transformer (ViT)-based foundation model across patient demographics with respect to sex, age, and ethnicity for the task of skin lesion classification using several public benchmark datasets. The significant advantage of this method is that conformal analysis is method independent, and it not only provides a coverage guarantee at the population level but also provides an uncertainty score for each individual. This is used to demonstrate the effectiveness of utilizing these embeddings for specialized tasks like diagnostic classification, meanwhile reducing computational costs. Secondly, the public benchmark datasets we used had severe class imbalance in terms of the number of samples in different classes. We used a model-agnostic dynamic F1-score-based sampling during model training, which helped to stabilize the class imbalance. We investigate the effects on uncertainty quantification (UQ) with or without this bias mitigation step. Thus, our results show how this can be used as a fairness metric to evaluate the robustness of the feature embeddings of the foundation model (Google DermFoundation), advancing the trustworthiness and fairness of clinical AI.

Abstract Biological and behavioral differences between genders influence infectious disease dynamics. Yet, most epidemiological models overlook these aspects in favor of age stratification alone. Here, we systematically evaluate the impact of incorporating gender-specific features into an age-structured epidemic compartmental model, calibrated to COVID-19 mortality data from the second wave in Italy (Autumn 2020–Winter 2021). We develop eight model versions representing different combinations of three data-driven features: gender-stratified contact matrices derived from CoMix data, gender-specific infection fatality ratios (IFR), and gender-dependent transmission rates linked to behavioral differences. We calibrate these models against aggregated mortality data and evaluate their performance on data disaggregated by gender, age, and both. Our results demonstrate that models incorporating gender-stratified contact patterns significantly outperform those relying solely on age, improving the accuracy of the fit even when analyzing age-disaggregated data alone. Furthermore, the inclusion of gender-specific IFR is essential for reproducing the empirically higher mortality rates observed in males. While phenomenological behavioral adjustments improve the fit for specific subgroups, such as older males, we observe trade-offs where maximizing performance for one demographic group occasionally reduces accuracy for another. Overall, our findings highlight that integrating gender data—particularly regarding contact patterns—is a critical step toward increasing the realism and precision of epidemiological models, even when outcome data is not fully disaggregated. Author summary Due to a combination of behavioral and biological factors, such as the adoption of protective measures and immune responses, infectious diseases affect men and women differently. However, the mathematical models used to understand and predict epidemics often overlook these aspects, focusing primarily on age. Here, we investigate whether explicitly including gender-specific data improves the accuracy of epidemic simulations. Using data from the second wave of the COVID-19 pandemic in Italy, we develop and compare several models that incorporated gender differences in social interaction patterns, biological risk of death, and protective behaviors. We find that men and women exhibit distinct social mixing patterns, and including this information is the most important factor for accurately reproducing the overall number of deaths. Additionally, accounting for the higher biological risk faced by males is essential for capturing gender-specific mortality trends. We also observe that because men had higher mortality rates, the models naturally prioritized fitting male data, sometimes at the expense of accuracy for females. Our work demonstrates that gender-agnostic models may miss crucial details. We conclude that incorporating gender-specific data is vital for creating more realistic models and designing inclusive public health policies that protect all demographic groups effectively.

ABSTRACT Challenges such as hallucination and non-determinism are well known to undermine the reliability of Generative AI (GenAI) applications. Despite this, their use across multiple fields and domains continues to rapidly grow. Many GenAI applications now generate or interact with geospatial information – leading to the potential proliferation of non-human generated geospatial data. Because such tools can be developed without specialist geographic expertise, there is a growing risk of diminished oversight and declining geospatial data quality. This paper presents an approach for evaluating the geospatial information produced by large language models (LLMs) using established frameworks for geospatial data quality. We apply this framework in two experiments that assess the completeness, accuracy, precision, and consistency of LLM outputs: geoparsing and route finding. Interpreting these results through ontological and teleological perspectives reveals fundamental limitations in how LLMs represent and reason about space. The study highlights the need for responsible evaluation and adaptation of data quality standards as GenAI becomes increasingly embedded within geographic information science.