Research Archive

Journal Article Open Access Green Computing

Carbon-Aware Computing at the Infrastructure Level: Dynamic Workload Migration, Renewable Energy Signal Integration, and Scope 3 Emission Attribution for Hyperscale Cloud Operators

Hyperscale cloud operators managing distributed data center fleets across multiple grid regions have a structural opportunity to reduce operational carbon emissions by dynamically migrating or deferring workloads to facilities with higher real-time renewable energy availability -- a practice termed carbon-aware computing. While carbon-aware scheduling at the application and DevOps pipeline levels has received growing research attention, the infrastructure-level engineering challenges of carbon-aware workload migration across hyperscale fleets -- including migration cost modeling, renewable signal integration latency, Scope 3 attribution methodology, and workload SLA compatibility analysis -- have not been comprehensively treated in the academic literature. This paper presents CarbonFleet, a carbon-aware infrastructure management system for hyperscale cloud operators, incorporating real-time carbon intensity signals from the Electricity Maps API across 34 grid regions, predictive renewable availability modeling using weather-driven solar and wind generation forecasts, and a multi-objective workload migration optimizer balancing carbon reduction, migration cost, and SLA compliance. CarbonFleet is evaluated in a simulation environment calibrated to a real 18-region cloud fleet over 12 months of historical demand and carbon intensity data. CarbonFleet achieves a 31.4 percent reduction in operational carbon intensity compared to carbon-unaware scheduling baselines while maintaining 99.94 percent SLA compliance across migrated workloads. We introduce a Scope 3 Cloud Emission Attribution Methodology (SCEAM) that disaggregates supply chain and customer workload carbon contributions with 94 percent fidelity against direct metering baselines, addressing a critical gap in cloud sustainability reporting standards.

Adunola Nwachukwu, Lars Svensson, Kenji Fujita, Amira El-Sayed· May 2026· 29 citations

Cite View PDF

Journal Article Open Access Artificial Intelligence

Multimodal Foundation Models for Scientific Discovery: Unified Architectures for Protein Structure Prediction, Materials Property Forecasting, and Drug-Target Interaction Modeling

The dominant paradigm in scientific AI has been domain-specific model development -- AlphaFold for protein structure, GNoME for materials discovery, and MolBERT for molecular property prediction -- each trained on domain-curated datasets with domain-specific architectural inductive biases. This paper challenges this paradigm by evaluating whether multimodal foundation models trained jointly across biological, chemical, and materials science modalities can achieve competitive or superior performance on domain-specific benchmarks while enabling cross-domain transfer capabilities unavailable to single-domain models. We present SciFoundation-7B, a 7-billion parameter multimodal scientific foundation model trained on a 2.4 trillion token corpus spanning protein sequences, crystal structures, SMILES molecular representations, scientific literature, and experimental assay data. SciFoundation-7B is evaluated on 14 benchmark tasks spanning protein structure prediction, drug-target interaction, toxicity prediction, band gap forecasting, and synthesis route planning. On 11 of 14 benchmarks, SciFoundation-7B matches or exceeds domain-specific state-of-the-art models within 2.1 percentage points, while uniquely enabling cross-domain transfer: protein-materials transfer learning improves metalloprotein binding affinity prediction by 12.4 percentage points versus single-domain baselines. We analyze the emergent cross-domain reasoning capabilities through systematic probing and introduce the Scientific Generalization Coefficient (SGC) as a measure of foundation model cross-domain utility. This work provides evidence for a potential paradigm shift toward unified scientific AI.

Obiageli Eze, Nina Lindqvist, Keiko Nakamura, Mohamed Mansour· Apr 2026· 38 citations

Cite View PDF

Journal Article Open Access Cybersecurity

Post-Quantum Cryptography Integration in DevSecOps Pipelines: Migration Strategies, Toolchain Readiness, and Organizational Preparedness for NIST PQC Standards

The finalization of NIST Post-Quantum Cryptography (PQC) standards in 2024 — including ML-KEM (Kyber), ML-DSA (Dilithium), and SLH-DSA (SPHINCS+) — has initiated an era of cryptographic migration urgency that places significant demands on DevSecOps pipelines, software supply chain tooling, and organizational security governance. This paper presents the first empirical study of PQC integration within DevSecOps contexts, examining organizational preparedness, toolchain readiness, and migration strategy effectiveness across 17 organizations in finance, defense, and critical infrastructure sectors. Using a combination of practitioner surveys (n=312), cryptographic inventory analysis, and migration pilot case studies, we characterize the PQC DevSecOps migration challenge along four dimensions: Cryptographic Asset Inventory Completeness, Pipeline Toolchain PQC Readiness, Certificate and Secret Rotation Automation Maturity, and Hybrid Classical-PQC Transition Governance. Our findings reveal that 78% of organizations lack comprehensive cryptographic asset inventories — the foundational prerequisite for migration planning — and that fewer than 12% of CI/CD toolchains natively support PQC signature verification. We introduce the PQC Migration Readiness Index (PMRI) and develop a phased PQC DevSecOps Migration Roadmap comprising five stages from cryptographic discovery through full PQC-native pipeline operation. Case study evidence from two pilot organizations demonstrates the feasibility of achieving PMRI Stage 3 within 18 months with focused engineering investment.

Obiageli Osu, Lars Magnusson, Kenji Mori, Ana Beatriz Lopes· Apr 2026· 41 citations

Cite View PDF

Journal Article Open Access Autonomous Systems

Multi-Agent Reinforcement Learning for Cooperative Drone Swarm Coordination: Decentralized Policy Learning, Communication Protocol Design, and Emergent Collective Behavior in Search and Rescue Operations

Autonomous drone swarms operating in search and rescue environments must coordinate complex collective behaviors -- area coverage, target localization, survivor extraction, and obstacle avoidance -- in GPS-denied, communication-constrained, and dynamically hazardous conditions that preclude centralized coordination architectures. Multi-agent reinforcement learning (MARL) offers a pathway to decentralized cooperative policy learning that is robust to agent failure and communication disruption, but the engineering of MARL systems that produce reliable cooperative behavior at swarm scales of 10 to 100 agents in realistic environments remains an open challenge. This paper presents SwarmRL, a MARL framework for cooperative drone swarm coordination evaluated in three progressively realistic environments: 2D grid simulation, 3D physics simulation using AirSim, and a real-world outdoor evaluation with a 12-drone swarm. SwarmRL employs a shared policy architecture with agent-specific observation encoders, a learned communication protocol using Graph Attention Networks for neighbor-to-neighbor message passing, and a centralized training with decentralized execution (CTDE) approach using QMIX as the cooperative value decomposition baseline. In the simulated search and rescue benchmark, SwarmRL achieves 94.2 percent target discovery rate within 600 seconds for a 20-agent swarm, outperforming both handcrafted coverage algorithms (78.3 percent) and independent RL agents without communication (67.1 percent). The learned communication protocol transmits 89 percent less data than broadcast communication while maintaining equivalent coordination quality. Real-world evaluation with 12 drones achieves 87.4 percent target discovery rate, demonstrating meaningful sim-to-real transfer.

Kelechi Okafor, Astrid Holm, Yuki Nishimura, Salma El-Amin· Feb 2026· 44 citations

Cite View PDF

Journal Article Open Access Artificial Intelligence

Autonomous DevOps: Toward Self-Healing, Self-Optimizing Delivery Pipelines Through Multi-Agent AI Orchestration

Autonomous DevOps represents the frontier of pipeline automation: delivery systems capable not merely of executing predefined workflows but of reasoning about system state, diagnosing anomalies, proposing remediations, and implementing approved changes without human intervention. This paper presents the theoretical foundations, architectural design, and empirical evaluation of an Autonomous DevOps system (AutoDevOps-1) built on multi-agent large language model orchestration. AutoDevOps-1 comprises four cooperating agents — Pipeline Analyst, Remediation Planner, Change Executor, and Safety Auditor — coordinated through a shared context protocol and governed by a formal safety constraint language we term Pipeline Safety Assertions (PSA). We evaluate AutoDevOps-1 in a production-equivalent staging environment replicating the infrastructure of a large financial institution, executing 4,200 simulated incident scenarios over six months. The system autonomously resolved 71.4% of incidents within defined safety constraints, escalating the remainder to human operators with structured context packages that reduced human resolution time by 58%. We conduct adversarial testing against 14 prompt injection and manipulation attack vectors, achieving containment in 92.9% of cases. We discuss the organizational, ethical, and liability implications of autonomous pipeline management and propose a governance framework for responsible deployment. This work establishes the empirical and conceptual foundations for the next generation of DevOps automation.

Adaobi Eze, Felix Zimmermann, Keiko Arai, Mohamed Al-Farsi· Feb 2026· 89 citations

Cite View PDF

Journal Article Open Access Quantum Computing

Variational Quantum Algorithms for Portfolio Optimization: Benchmarking QAOA and VQE Against Classical Solvers on Near-Term Quantum Hardware

Variational quantum algorithms -- particularly the Quantum Approximate Optimization Algorithm (QAOA) and Variational Quantum Eigensolver (VQE) -- have been widely proposed as near-term candidates for achieving quantum advantage in combinatorial optimization problems relevant to finance, including portfolio optimization, risk parity, and credit risk assessment. However, realistic benchmarking of these algorithms on current quantum hardware against state-of-the-art classical solvers has been limited by hardware access constraints and lack of standardized evaluation methodologies. This paper presents the most comprehensive empirical benchmarking study of QAOA and VQE for financial portfolio optimization to date, executing experiments on IBM Quantum Eagle (127-qubit), IBM Quantum Osprey (433-qubit), and IonQ Aria (25-qubit) processors across portfolio sizes from 10 to 50 assets, and comparing solution quality and runtime against Gurobi, D-Wave Advantage (quantum annealing), and simulated annealing baselines. Current QAOA implementations at p=3 circuit depth achieve mean solution quality 18.4% below Gurobi on 30-asset portfolios, with hardware noise as the dominant limiting factor. D-Wave Advantage outperforms gate-model QAOA for portfolio sizes below 40 assets on current hardware. We introduce a Quantum Finance Readiness Index (QFRI) that maps algorithm performance trajectories against projected hardware improvement curves, estimating practical quantum advantage emergence for 100-asset portfolios at approximately 2028-2031 under optimistic noise reduction scenarios. This work provides financial institutions with an honest evidence base for quantum computing investment planning.

Chukwuebuka Aneke, Astrid Svensson, Daisuke Nishimura, Laila Benali· Aug 2025· 89 citations

Cite View PDF

Journal Article Open Access Software Engineering

Observability 2.0: Continuous Profiling, eBPF-Based Telemetry, and the Emerging Observability Stack for Cloud-Native DevOps at Scale

The three-pillar observability model — logs, metrics, and distributed traces — which has guided cloud-native monitoring strategy for nearly a decade, is showing structural limitations in the face of highly dynamic, ephemeral, and high-cardinality workloads characteristic of modern Kubernetes-native systems. This paper examines the emergence of continuous profiling as the "fourth pillar" of observability and evaluates the broader shift toward eBPF-based kernel-level telemetry as a foundation for low-overhead, always-on instrumentation. We present an empirical evaluation of continuous profiling tools (Pyroscope, Parca, and Grafana Pyroscope) and eBPF-based observability frameworks (Cilium, Tetragon, and Pixie) in a production-equivalent Kubernetes environment running a 120-service microservices benchmark. Continuous profiling reduced mean time to identify CPU and memory regression root causes from 47 minutes to 9 minutes compared to metric-only approaches. eBPF-based network telemetry eliminated the 12–18% performance overhead of sidecar-based service mesh observability while providing equivalent visibility. We introduce the Observability Stack Completeness Score (OSCS) and map 23 production organizations against it, finding that fewer than 15% have adopted continuous profiling despite its measurable impact. We propose the Observability 2.0 Reference Architecture and provide migration guidance from traditional three-pillar stacks.

Kelechi Nwachukwu, Astrid Bergqvist, Ryosuke Kawamoto, Diana Fonseca· Aug 2025· 98 citations

Cite View PDF

Journal Article Open Access Network Engineering

Intent-Based Networking with Large Language Models: Natural Language Network Policy Specification, Automated Configuration Generation, and Closed-Loop Verification in Enterprise Campus Networks

Intent-Based Networking (IBN) -- which allows network operators to specify desired network behavior in high-level business terms and rely on automation to translate intent into device configurations -- has been a long-standing aspiration of network management research. The emergence of large language models with strong natural language understanding and code generation capabilities has created new opportunities for practical IBN implementation without the manual ontology engineering that constrained earlier approaches. This paper presents NetIntent, an LLM-powered IBN framework that translates natural language network policy statements into vendor-agnostic network configuration artifacts (Ansible playbooks, Cisco IOS-XE, Juniper JunOS) with closed-loop verification against the deployed network state. NetIntent employs a three-stage pipeline: intent parsing and formalization using GPT-4o with network domain-adapted prompting, configuration synthesis using retrieval-augmented generation over a vendor configuration knowledge base, and automated verification using network simulation and configuration diffing. Evaluation is conducted across a 200-device enterprise campus network testbed spanning switching, routing, security policy, and QoS configuration domains. NetIntent correctly translates 87.4 percent of natural language policy statements into syntactically and semantically correct configurations, with the remaining 12.6 percent requiring human clarification due to intent ambiguity. Mean configuration generation time of 23 seconds compares favorably to a baseline of 47 minutes for manual configuration by experienced network engineers. We analyze the failure modes and introduce NetIntent-Guard, a post-generation policy correctness verifier reducing misconfiguration escape rate by 94 percent.

Chidiebere Eze, Maja Karlsson, Takeshi Yamamoto, Nadia Benali· Apr 2025· 98 citations

Cite View PDF

Journal Article Open Access Artificial Intelligence

AI-Driven Infrastructure Provisioning: Evaluating Natural Language to Infrastructure-as-Code Generation Using Large Language Models in DevOps Workflows

The application of large language models to Infrastructure as Code generation — enabling engineers to describe desired infrastructure in natural language and receive executable IaC manifests — represents a potentially transformative capability for DevOps productivity, but also introduces novel correctness, security, and auditability risks that have not been systematically evaluated. This paper presents the first rigorous empirical evaluation of natural language to IaC generation using GPT-4 Turbo, Claude 3.5 Sonnet, and Gemini 1.5 Pro across a benchmark suite of 840 infrastructure provisioning tasks spanning AWS, Azure, and GCP in both Terraform HCL and Pulumi TypeScript. We evaluate generated configurations across five quality dimensions: syntactic correctness, semantic correctness (does the configuration provision what was requested?), security policy compliance (evaluated against CIS Benchmarks), idempotency, and cost efficiency. Across all models and providers, syntactic correctness averages 91.4%, but semantic correctness drops to 67.8%, with the most significant gaps occurring in complex networking, IAM policy generation, and multi-region configurations. Security policy compliance averages 54.2%, with IAM over-permissioning as the most prevalent deficiency. We introduce an LLM-IaC Quality Score (LIQS) aggregating all five dimensions and propose an LLM-Assisted IaC Workflow incorporating human review gates, automated compliance scanning, and policy-as-code validation to compensate for identified deficiencies. This work establishes the empirical baseline for safe LLM-assisted infrastructure provisioning in DevOps.

Ebuka Okonkwo, Hanna Björk, Daisuke Fujiwara, Rita Cardoso· Apr 2025· 142 citations

Cite View PDF