Glaciers produce abundant grains of reactive minerals that influence inorganic carbon cycling, yet the links between mineral weathering and CO2 dynamics in glacier-fed streams (GFSs) remain poorly understood. Here we combine streamwater geochemistry, solute end-member modelling, carbonate saturation calculations and suspended-sediment mineralogy across 170 GFSs worldwide. We show that 69% of GFSs were undersaturated with respect to atmospheric CO2 equilibrium, indicating widespread atmospheric CO2 uptake. Grain-scale analyses of selected carbonate-bearing GFSs reveal metastable Mg-bearing carbonate phases consistent with secondary mineral formation in transient microenvironments. We propose that CO2 released during sulfuric acid-driven carbonate weathering and secondary calcite precipitation can be partly consumed by carbonation of silicate- and clay-derived mineral surfaces, transferring dissolved inorganic carbon into particulate inorganic carbon. This sediment-mediated pathway can contribute to CO2 undersaturation under favourable lithological and geochemical conditions, revealing a previously underappreciated coupling between chemical weathering, sediment–water interactions and inorganic carbon storage in proglacial landscapes rapidly expanding under global glacier retreat.

Distributionally robust optimization (DRO) minimizes worst-case risk under distributional uncertainty, but its decisions can be suboptimal in terms of their excess risk, or regret. We study this gap for multivariate least-squares regression, assuming that the observed source law and unknown target law differ by an adversarial 2-Wasserstein perturbation of size ε. We first identify the asymptotic complexity of robust decision-making in the ε→0 limit, characterized by a problem-specific condition number κ. Here, a wide spectrum of simple algorithms, including DRO, achieve the instance-optimal rate of O(κε^2). For fixed ε>0, we prove that DRO still achieves the minimax rate if the problem has bounded residual effective rank. On the other hand, we identify significant failure modes where DRO is provably suboptimal by dimension-dependent factors. Fortunately, we prove that two alternative procedures, minimax regret optimization (MRO) and spectral-norm regularization (SNR), attain the minimax rate. We derive an exact convex reformulation of MRO and an accelerated proximal algorithm for SNR, along with high-probability finite-sample guarantees. We support these theoretical results with numerical experiments separating the performance of MRO and SNR from DRO.

This document is a thematic taxonomy of the metrics and key performance indicators (KPIs) used, or proposed, to describe datacenters from chip to site and beyond the site boundary. It lists 619 metrics and measurements and 266 candidate KPIs in 28 families, grouped into seven sections: energy and electrical systems; thermal management, water and environment; IT systems and workloads; sustainability, lifecycle and integrated performance; reliability, resilience and service; capacity, operations and economics; and safety, security and governance. Each entry has a stable identifier and a definition; metrics carry a unit, KPIs a formula written with metric identifiers and a direction. Every entry also records its origin, a verification status against primary or secondary sources, a maturity class (117 standardized, 647 in use, 121 proposed), its measurement level from chip to organisation, and dimension flags for sustainability, efficiency, security, health and safety, and economics. A listed KPI is a candidate: it becomes a KPI only with an objective, a target and an owner. Cooling is split into the facility plant, the air path and direct liquid cooling, with KPIs that compare the liquid path with the whole site. calculation pipelines and a gaming-risk checklist, and draws the catalogue as hypergraphs in which families, sections, dimension flags and measurement levels are hyperedges and identifier references are directed edges; hovering a node shows the definition of the entry. Abbreviations are defined under every table. The L A T E X source and the figures are generated by Python scripts distributed with the document.

This document is a hierarchical taxonomy of the technical elements of a datacenter, from the site and the building down to the chip and the software that runs on it. It decomposes a datacenter into 11 domains, 78 subsystems, 358 equipment classes and 378 components or variants. Every node has a stable identifier and a definition; equipment classes carry indicative engineering parameters, applicable standards (171 documents named) and typical failure modes. Domain boundaries follow the design and operational parts of EN 50600. Every node carries type facets (criticality, energy role, resource flows, EN 50600 part, ISO/IEC 30134 KPI, and the measurement level on the scale of the companion catalogue, from L0 chip to L6 site), drawn as one label system shared by the entries, the figures and the JSON export. The taxonomy is linked to the companion catalogue Datacenter Metrics and Key Performance Indicators: 141 nodes point to 527 of its 885 metrics and candidate KPIs, in 33 of its families and sub-families, so that each piece of equipment can be traced to what is measured at it and each indicator to the equipment it measures. The taxonomy is drawn as hypergraphs in the conventions of the catalogue (an overview by EN 50600 part, one map per domain, resource flows in extra-node form, measurement levels in incidence form), and both documents are joined in one hyper-bag-graph in which catalogue families are hyperedges over taxonomy domains with multiplicities; hovering a node shows its definition. Equipment classes connect through 968 ports standardised on 71 media, grouped in the twenty interface families of the parallel taxonomy v0.7.0, whose engineering sheets (boundaries, quantities with units, named parts) and representative assemblies are imported where they map, and whose 121 families are cross-walked onto this taxonomy; interfaces, physical domains and media in and out are exposed as facets for visualisation. The 40 bibliographic entries record what was verified and where; parameters, failure modes, facets and links are marked as not verified. The LaTeX source, the figures and the JSON export are generated by Python scripts from YAML data distributed with the document.

This conversation with architect and professor Sophie Delhay is part of a broader discussion on the project of housing that is ongoing at Burning Farm. Like the previous practices we interviewed—Lacol, SUMMACUMFEMMER, Plan Común—Delhay has produced some of the most experimental housing projects of the last decade.
Our interview with Sophie Delhay offers insight into her approach to housing design, including the intricate relationship between private and collective spaces within housing and the constant tension between variation and repetition that characterizes her approach to designing large housing. Delhay explores the tension that arises when an architect designs housing between the desire for clear architectural forms and the need to allow architecture to fade into the background. In our interview, we discussed the current state of social housing, client expectations, and market pressures, in an effort to question the relationship between research and practice.

In this thesis, we present the advances in self-organised nano-pillar lattice fabrication, including improvements in graphene (Gr) transfer onto targets in ultra-high vacuum (UHV), demonstrated increase of Gr capped self-organised cluster thermal stability, and creation of a high-density nano-pillar heterostructure achieved by second layer self-organised cluster growth on a transferred 2D material patterned by the clusters below it. We first present an ion intercalation-based graphene delamination method for the creation of a PTFE-supported graphene bilayer from which one monolayer is transferred onto the target sample using wafer bonding in UHV. With this method we demonstrate mm-scale graphene coverage onto Ir(111) and Irclust/Gr/Ir(111) targets in UHV. The novel method was tested alongside the acid etching-based bilayer preparation method and compared using Raman spectroscopy. The ion intercalation delamination method was found to result in higher quality and homogeneity bilayer graphene stamps, compatibility with graphene growth on reusable Pt substrates with larger grains further increasing graphene quality, as well as lower doping due to avoiding exposure to etchants. After transfer on Ir(111), scanning tunnelling microscopy (STM) imaging revealed significantly larger domain sizes of the monolayer Gr transferred. Additionally, Auger electron spectroscopy (AES) measurements were used to quantify graphene transfer coverage on a large scale, showing around 50% coverage over mm scale for both Ir(111) and Irclust/Gr/Ir(111) targets. Then we investigate the effects on thermal stability of Gr-capped Irclust/Gr/Ir(111) in UHV. STM imaging of samples prepared at varied temperatures revealed increased resistance against Ostwald ripening and peripheral diffusion, with filling factor values 20% - 30% higher in Gr-capped areas areas for a temperature range from 750 K to 1250 K, demonstrating the usefulness of the system for magnetic nano-pillars with high blocking temperature. Finally, the fabrication of a high-density self-organised nano-pillar heterostructure was demonstrated by subsequent cluster growth, Gr transfer, and growth of a second layer of Ir on Gr/Ir(111). The resulting Irclust/Grtransfer/Irclust/Gr/Ir(111) heterostructure has been investigated using STM, with second layer clusters shown to self-organise on top of Gr-capped clusters, and apparent height measurements demonstrating two Ir-layer high clusters grown on top of transferred graphene. An alternative system that could potentially be extended to 3D growth of self-organised Co clusters on Au(11,12,12) surface has been capped with Gr on the mm-scale as observed using AES, with capped areas showing a preserved self-organised pattern as investigated with STM. The demonstrated achievements offer the foundations for further development of nano-pillar heterostructure with the potential for high-density magnetic nano-pillars.

Optical frequency combs link optical frequencies to electronically countable microwave sig- nals. Kerr microresonators miniaturize these sources, but extending their bandwidth remains challenging. The central aim of this thesis is to explore routes to broader optical frequency combs through the joint engineering of dispersion, resonator geometry, and excitation in three integrated architectures. The first part combines pulsed pumping and dispersion engineering in a silicon nitride (Si3N4) Euler racetrack. Synchronous electro-optic pulses supply high peak power at moderate average power. Adiabatic bends suppress parasitic spatial-mode conversion, while dispersion engineering and broadband extraction support an octave-spanning soliton microcomb at an electronically detectable repetition rate. Radio-frequency and heterodyne measurements distinguish the soliton from modulation instability and verify coherence at widely separated wavelengths. The second part combines gallium phosphide with photonic-crystal Fabryâ Pérot cavities whose chirped reflectors supply anomalous round-trip dispersion. Picosecond subharmonic pumping overcomes the thermo-refractive access barrier, enabling the first room-temperature dissipative Kerr solitons in GaP. Heterodyne noise measurements reveal the transfer of timing noise from the electro-optic drive. A modified Lugiatoâ Lefever model separates the nonlinear detuning shift caused by counter-propagating fields from the period-four dynamics and interleaved subcombs imposed by the drive. Finally, the Vernier effect in coupled rings reshapes the resonance spectrum into recurring hybridized dispersion valleys. A transfer-matrix model guides two device generations and relates measured hybridization to broadband comb generation. The analysis identifies a novel four-wave-mixing mechanism involving intra- and interband photon transfer between these valleys, offering a new route to spectral expansion in coupled-resonator systems. Together, these studies show that useful broadband microcombs require joint control of dispersion, loss, coupling, thermal dynamics, and the drive. Future work will focus on demon- strating f â 2f self-referencing, integrating pump-pulse generation and compression on chip, and achieving phase-locked comb generation across multiple Vernier valleys.

Membrane-based gas separation offers an energy-efficient alternative to conventional separation processes, but practical membranes must combine high permeance, selectivity, stability, mechanical robustness, and scalable fabrication. Carbon materials are promising candidates for gas separation membranes because of their structural versatility and chemical and physical stability. When fabricated as ultrathin selective layers with well-controlled molecular-sieving pores or channels, carbon membranes can combine short transport pathways with strong size selectivity. This thesis focuses on the design, synthesis, and characterization of ultrathin carbon-based membranes for gas separation, aiming to control molecular-sieving pores and channels at the angstrom scale to discriminate gas molecules with small differences in kinetic diameter. Two representative transport geometries are investigated: through-plane nanopores in porous graphene and confined channels between carbon layers or domains. The first part studies bottom-up synthesis of porous nanocrystalline graphene as an ultrathin molecular-sieving membrane. Nanopores are incorporated during graphene growth rather than by post-synthetic etching. By controlling the cooling rate during precipitation-based growth on nickel, graphene grain intergrowth and effective pore size can be tuned. Rapid cooling reduces the effective pore size and improves H2/CH4 selectivity from 6.2 to 53.3, demonstrating a synthesis-based strategy for angstrom-scale pore tuning. The second part develops 10-nm-thick carbon membranes derived from poly(4-vinylpyridine) by pyrolysis under controlled oxygen exposure. Trace O2 modifies the decomposition pathway of the polymer precursor during pyrolysis and transforms the resulting carbon membrane from a relatively uniform network into a disrupted ultramicroporous structure. The resulting multimodal ultramicroporous channels provide rapid transport pathways while retaining molecular-sieving capability. These 10-nm-thick carbon membranes achieved H2 permeance above 10,000 GPU (gas permeation units, 1 GPU = 3.35 × 10 -10 mol m-2 s-1 Pa-1) and H2/N2 mixture selectivity above 200 at 150 °C, showing that oxygen-modulated pyrolysis can overcome the thickness limitation of conventional carbon molecular sieve membranes. The third part addresses the challenge of structural characterization in fragile porous graphene membranes. Reliable pore analysis requires clean, suspended, and mechanically stable specimens, but conventional polymer-assisted transfer can leave residues that obscure nanopores, while direct transfer can damage porous films. Therefore, a polymer-free carbon nanotube mesh scaffold, termed CMesh, is developed to mechanically reinforce porous graphene during transfer and imaging. This scaffold enables robust atomic-resolution TEM/STEM characterization of pre-existing porous graphene and in situ heating studies up to 900 °C, providing a practical platform for linking nanopore structure with membrane transport. Overall, this thesis demonstrates an integrated strategy for engineering ultrathin carbon-based membranes by connecting synthesis, angstrom-scale pore/channel control, gas separation performance, and atomic-resolution characterization. The developed approaches provide new routes to tune through-plane nanopores and confined ultramicroporous channels, contributing to the development of high-performance molecular-sieving membranes for energy-efficient gas separation.

Coral reefs are among the most biodiverse and most threatened ecosystems on Earth, declining under ocean warming and recurrent mass bleaching. Whether they persist depends on how fast corals can adapt. In this context, this thesis assesses the adaptive potential of the reef-building coral Acropora millepora in the Coral Sea from a holobiont perspective, reading the environmental seascape, the host genome, the algal symbionts and the colony phenotype together across fourteen reef sites in two contrasting regions: the Capricorn Bunker Group, at the southern tip of the Australian Great Barrier Reef, and the remote Chesterfield-Bellona plateau of New Caledonia.

The environmental seascape was characterised from satellite remote sensing, as eighty descriptors of thermal stress, water quality, salinity and currents computed over four time windows, and was used to guide sampling across the environmental range rather than across geography alone. Colony morphology, colour, health and microhabitat were scored from standardised in situ photographs of 268 colonies. The host genome of 236 colonies was resolved by low-coverage whole genome sequencing (lcWGS), and the algal symbionts of 237 colonies by ITS2 metabarcoding. Genome-environment associations were then scanned with LFMM2 and SAMbada, cross-checked against the environment-free PCAdapt selection scan, and integrated, alongside the symbionts and the phenotype, in redundancy analyses that projected a polygenic adaptive index across the reefs.

The two regions form two genetically distinct units, each effectively panmictic within itself, separated by low but consistent differentiation and with Chesterfield-Bellona carrying about three times the nucleotide diversity of the Capricorn Bunker Group. Within Chesterfield-Bellona, genetic differentiation increases with geographic distance, a stepping-stone pattern absent from the Capricorn. Phenotype, symbiont community and host genome separate the same two regions, the dominant Cladocopium type profile switching between them. Candidate loci associated with sea surface temperature anomalies and heat-stress metrics fall in molecular chaperones, an antioxidant transporter, and extracellular-matrix and cilia genes, a set that recurs in independent global coral studies. The adaptive index draws a warm, clear-water-adapted pole across the Chesterfield-Bellona plateau, strongest in its north, grading to a cooler, more turbid-adapted pole in the Capricorn Bunker Group. It is on the Chesterfield-Bellona plateau that the adaptive signal of the host, the symbiont community and colony health agree, a first step towards seascape genomics at the scale of the holobiont.

Because the two regions lie far apart, the regional contrast between them is confounded with the environment, the symbionts and the host structure, so adaptive signals within each region were harder to resolve. However, sampled in June and September 2023, months before the severe 2024 El Niño marine heatwave, the colonies offer a rare opportunity to test in the field whether the genotypes predicted to be heat-adapted actually survived. Conserving the adaptive potential of corals, in a climate that no longer holds still, means reading the host, the symbiont and the environment together.

AI control is a field of study intended to provide a second layer of defense should our efforts at AI alignment fail. Starting with the premise that an AI system is misaligned, AI control is concerned with designing mechanisms to either prevent misbehavior or catch the AI red-handed. In this thesis, we present an overview of AI control alongside several contributions to the field, split into the following chapters:

Overview of AI control. To motivate AI control, we begin with a survey of theoretical and empirical evidence of misalignment in modern AI systems. Then, we discuss the high-level principles that guide AI control mitigations, covering common deployment contexts, the specific threats that AI control aims to prevent, and how AI control evaluations are conducted. We also introduce a distinction between concentrated and diffuse threats from untrusted AI models and discuss why these two types of threats demand different mitigations. Concentrated threats are those that unfold over a short period of time. The science of concentrated threats, sometimes called concentrated AI control, is more developed; following prior work, we present a mathematical model that can be used to evaluate the safety of concentrated AI control protocols.

Control tax. In this chapter, we study a concentrated AI control protocol called trusted monitoring. Trusted monitoring proceeds by having a weaker trusted AI model assign suspiciousness scores to the actions of the untrusted model, with the most suspicious actions escalated for human auditing. To evaluate trusted monitoring, we have language models generate solutions to competitive programming problems that contain subtle bugs. We then test how well other models can distinguish these "backdoored" solutions from the correct ones. Using the mathematical model from the previous chapter, we show how to compute the safety of trusted monitoring analytically in the limit of a large number of deployed AI agents. Finally, we combine our applied and theoretical results under the framework of the control tax---the operational and financial cost of integrating trusted monitoring into AI pipelines.

Adaptive attacks on AI control protocols. In this chapter, we present simple attacks that subvert existing monitoring-based protocols in concentrated AI control. The most common kind of trusted monitor used in control protocols is a prompted language model, and language models are notoriously susceptible to prompt injections. We exploit this vulnerability to demonstrate that trusted monitors can be induced to label most of the code they see as not suspicious, rendering them essentially useless. Our language model-generated prompt injections generalize across several monitors. Using the mathematical model from the previous chapters, we show the disastrous downstream effects of these prompt injections on the safety of multiple control protocols.

Diffuse AI control on fuzzy tasks. In this chapter, we propose a red-teaming framework for evaluating training interventions against diffuse threats from a misaligned AI, such as deliberately underperforming on difficult-to-grade tasks. We show that a misaligned AI can subvert the process of generating ML experiment ideas even when optimized against scores assigned by a weaker model. We also show that there exist prompts for the weak scorer that significantly improve its robustness, though efficiently discovering such prompts with realistic affordances remains an open problem.