Scientific papers

Research paperPDF

Journal of Physics: Conference Series (IOP Publishing) · 2018

Evaluation of Oxygen Reduction System (ORS) in Large-scale Fire Tests

Compare ORS design concentration with the 15-16% standard-test values in VdS 3527 and EN 16750; since large-scale hard-limit LOCs (11.1% for cartoned commodities, 13.0% for uncartoned plastics) are lower, justify the design safety margin against them.

• Earlier lab work found LOCs of 10-13% for cardboard, wood, PS and PE, below the 15-16% range from standard tests in VdS 3527 and EN 16750:2017.

Fire safetySecurity

Read / Download

Research paperPDF

Buildings (MDPI AG) · 2024

Experimental and Numerical Investigation of Airflow Organization in Modular Data Centres Utilizing Floor Grid Air Supply

Choose raised-floor height of 600-800 mm and grille open area of about 40%; a 1000 mm floor gave only a small extra gain, and 20% or 60% open area pushed cabinet intake above 27 °C.

• With cold-aisle containment entransy loss was 4450 kW·K, and with both cold and hot aisles contained it was 3500 kW·K.

AirflowCooling

Read / Download

Research paperPDF

Batteries (MDPI AG) · 2025

Experimental and Reduced-Order Modeling Research of Thermal Runaway Propagation in 100 Ah Lithium Iron Phosphate Battery Module

In LFP module design, evaluate inter-cell thermal resistance (interface insulation) together with measures that cut heat release to stop propagation; in the model an interface resistance of hR = 100 W·m^-2·K^-1 alone suppressed it.

• Venting phases (vent opening to runaway) lasted 204 s, 399 s, 156 s and 309 s for cells 1-4.

Fire safetyUPSReliability

Read / Download

Research paperPDF

Batteries (MDPI AG) · 2026

Experimental Study on Fire Suppression of Lithium-Ion Battery Module with Different Extinguishing Agents in Confined Space

For confined battery packs and cabinets, evaluate a fast flame-knockdown agent (such as heptafluoropropane) together with an agent or approach giving long-lasting cooling (such as perfluorohexanone); no single agent stopped heat transfer to neighbouring cells.

• Without suppression, peak HRR reached 69.09 kW and total heat release was 8.05 MJ, with multiple deflagrations after venting in the confined space.

Fire safetyUPS

Read / Download

Research paperFree to read

Proceedings of the AAAI Conference on Artificial Intelligence (Association for the Advancement of Artificial Intelligence (AAAI)) · 2025

Fast 3D Surrogate Modeling for Data Center Thermal Management

Offer a fast-preview thermal mode with a voxel surrogate trained on the tool's own CFD runs: inputs of rack power by segment, CRAH setpoints, fan speeds and vent locations; output a 3D inlet-temperature map in under a second for layout iteration.

• Surrogate inference took about 100 ms against about 35 minutes for CFD; the abstract states about a 20,000x speed-up.

AirflowCoolingOperations

Research paperPDF

IEEE Access (Institute of Electrical and Electronics Engineers (IEEE)) · 2025

Fast Frequency Response Potential of Data Centers Through Workload Modulation and UPS Coordination

For a UPS that will provide grid services, define the SOC window (for example 20-80%) and show by calculation that backup autonomy is preserved inside it.

• The coordinated strategy cut nadir deviation by 57% and recovery time by 34%, with task delay below 2%.

UPSElectricalEnergy

Read / Download

Research paperPDF

Sustainability (MDPI AG) · 2024

Feasibility Analysis of Indirect Evaporative Cooling System Assisted by Liquid Desiccant for Data Centers in Hot-Humid Regions

Where plain IEC falls short in a hot-humid climate (high wet-bulb), consider a liquid-desiccant pre-stage; in the Hong Kong example supplementary mechanical cooling fell from 5092 to 31 h.

• With a 25 °C supply-air threshold, supplementary mechanical cooling hours fell from 5092 h to 31 h per year.

CoolingEnergy

Read / Download

Research paperPDF

Energies (MDPI AG) · 2023

Free Cooling for Saving Energy: Technical Market Analysis of Dry, Wet, and Hybrid Cooling Based on Manufacturer Data

Check free-cooling feasibility from the required coolant outlet temperature, the tower approach (dry 10 K, open wet 5 K, closed wet 6.5 K) and the at-least-5 K coolant range.

• Minimum outlet temperatures with evaporative cooling are up to 16 K lower than with dry cooling.

CoolingWaterEnergy

Read / Download

Research paperPDF

arXiv (Aalborg University) · 2026

From Grid to Chip: Power Architecture, Stability, and Flexibility of AI Data Centers

For large sites, capture grid-code inputs per jurisdiction (FRT curve, MW/min ramp limit, power-quality limits) as project inputs and flag UPS settings that transfer to battery or generator on shallow sags.

• NVIDIA reports that 800 V DC instead of 415 V AC cuts copper requirements by 45%.

ElectricalUPSReliability

Read / Download

Research paperFree to read

IET Generation, Transmission & Distribution (Institution of Engineering and Technology (IET)) · 2026

From Liability to Asset: A Three-Mode Grid-Forming Control Framework for Centralized Data Center UPS Systems

For large or AI facilities, record the UPS's fault ride-through behaviour (momentary cessation versus ride-through with minimum grid draw), the inverter current limit and the post-fault ramp rate, alongside the point-of-interconnection short-circuit ratio.

• Unserved IT energy was 0.00000 MWh with the proposed controller versus 0.00208 MWh for a momentary-cessation benchmark.

UPSElectricalReliability

Research paperPDF

arXiv preprint (Stanford University / Microsoft Azure Research) · 2026

From Servers to Sites: Compositional Power Trace Generation of LLM Inference for Infrastructure Planning

For an AI inference facility, set interconnection capacity from traces generated from the workload, not nameplate TDP or mean power; TDP sizing gave 1.19 MW against 0.75 MW trace-based (about 60% too much).

• Generated facility traces show a peak-to-average ratio of 1.19, a load factor of 0.84 and a maximum 15-minute ramp of 0.11 MW.

EnergyElectricalRacks

Read / Download

Research paperPDF

Energy Advances (Royal Society of Chemistry (RSC)) · 2025

Full life cycle assessment of an industrial lead–acid battery based on primary data

For UPS duty with a few cycles a year and minutes of autonomy, compare battery chemistries by environmental impact categories (GWP, ecotoxicity, resource depletion) as well as energy density, because production and EOL dominate.

• Data centres account for about 37% of the roughly 3.4 billion USD global stationary-battery market, followed by telecom at 25%.

UPSSustainability

Read / Download

Research paperPDF

IEEE Access (Institute of Electrical and Electronics Engineers (IEEE)) · 2026

GRAPE-UPS: Direction-Constrained Peer Anomaly Detection for Weak Batteries in Data-Center UPS Strings

Require block-level (per-unit) voltage and resistance monitoring in UPS battery monitoring rather than string totals, citing that one weak unit in a series string limits the whole string's capacity.

• On an independent electrochemical simulation benchmark, directed peer scoring achieved AUPRC 0.848 under nominal conditions and 0.850-0.877 under operating shifts.

UPSOperationsReliability

Read / Download

Research paperPDF

arXiv preprint (University of Michigan) · 2026

Grid Integration of AI Data Centers: A Critical Review of Energy Storage Solutions (arXiv preprint)

Layer storage by timescale (chip/server sub-ms to ms, rack ms to seconds, UPS and GiUPS sub-second to 10 s, BESS minutes to hours) and show each layer's duty in the coordination scheme.

• Rapid AI load variations often induce 5-10% voltage swings within milliseconds.

UPSElectricalEnergy

Read / Download

Research paperPDF

Advances in Applied Energy (Elsevier BV) · 2026

Grid Integration of AI Data Centers: A Critical Review of Energy Storage Solutions

Layer storage by timescale: GPU/server supercapacitors for sub-second swings, rack BBUs for minutes, facility GiUPS and BESS for grid services and longer durations; show each layer's duty and coordination separately in the architecture.

• BESS cost fell from about $7,500/kWh in 1991 to $133/kWh in 2024, with about $80/kWh projected by 2030.

UPSElectricalEnergy

Read / Download

Research paperPDF

Energies (MDPI AG) · 2020

Highly Renewable District Heat for Espoo Utilizing Waste Heat Sources

In selling data-centre waste heat, set heat-pump COP by cooling type (all-liquid 6.3, mixed 5.5, air-cooled 4.1) and include electricity price sensitivity in the feasibility study.

• The Espoo DH network serves over 80% of the population in Espoo and neighbouring cities, with total annual heat demand of about 2.6 TWh.

SustainabilityEnergy

Read / Download

Research paperFree to read

arXiv preprint (Texas A&M University) · 2026

How AI Load Swings Reduce Datacenter Power Conversion Efficiency

In power-chain loss calculators for AI halls, do not evaluate PSU, UPS and transformer efficiency only at average load; accept a load-profile descriptor (duty cycle, high/low power, frequency).

• Cycle efficiency is computed as (D*PH + (1-D)*PL) / (D*PH/etaH + (1-D)*PL/etaL).

ElectricalEnergyRacks

Research paperFree to read

Leibniz Supercomputing Centre · 2025

HPC and AI in Supercomputing Centres: Green Today, Greener Tomorrow (SCA25 Workshop on Energy Efficiency, LRZ)

Let liquid-cooled design templates allow facility supply temperatures up to 45 C with dry-cooler-only heat rejection, and show the resulting PUE.

• Systems span from SuperMUC (2012) to SuperMUC-NG Phase 2.

Liquid coolingEnergySustainability

Research paperPDF

Applied Thermal Engineering (Elsevier BV) · 2025

Immersion cooling fluid comparison using a single-phase immersion/liquid technique for data centers

In hybrid immersion/direct liquid cooling design, rank dielectric fluids by FOM1, but remember a high-FOM1 fluid can infiltrate the TIM of cold-plated CPUs and raise their temperature.

• A near doubling of FOM1 lowered average fluid temperature by approximately 7.7 °C.

Liquid coolingRacksReliability

Read / Download

Research paperPDF

Buildings (MDPI AG) · 2024

Improving Thermal Performance in Data Centers Based on Numerical Simulations

If cabinets in the rows nearest the cooling units run hot (high-velocity plenum flow starving the nearest tiles), consider adjustable underfloor deflectors and grille opening adjustment before fan-assisted tiles (18.1% and 4.9% airflow shift).

• The largest drop in average cabinet outlet temperature was 2.81 °C.

AirflowCooling

Read / Download

Research paperPDF

Energies (MDPI AG) · 2019

Investigation of Heat Management in High Thermal Density Communication Cabinet by a Rear Door Liquid Cooling System

Consider a cabinet-level rear-door heat exchanger as an option that can raise heat dissipation by 56-67% over air cooling; compare oil or water coolant separately on leakage risk and maintenance.

• For simulated heat-source power of 800-1600 W, average heat-dissipation efficiency of the RDHE was 44.1%.

Liquid coolingRacksCooling

Read / Download

Research paperFree to read

Proceedings of the 7th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (ACM) · 2020

Kalibre: Knowledge-based Neural Surrogate Model Calibration for Data Center Digital Twins

Include a calibration workflow in the hand-over to operations: import aisle sensor data, mark per-server airflow (and bypass/recirculation) as calibration parameters with priors, and report MAE before and after calibration.

• DOI 10.1145/3408308.3427982; arXiv:2001.10681.

AirflowCommissioningOperations

Research paperFree to read

Proceedings of the 22nd International Middleware Conference (ACM) · 2021

Let's Wait Awhile: How Temporal Workload Shifting Can Reduce Carbon Emissions in the Cloud

For carbon and sustainability reporting, use hourly (not annual average) grid carbon-intensity profiles per region.

• Batch jobs make up 60% of processing on large clusters, and more than 40% of these run daily.

SustainabilityEnergyOperations

Research paperPDF

International Conference on Learning Representations (ICLR 2024) · 2024

LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models

In AI data-centre carbon estimation, compute operational (FLOPs, power, PUE, grid intensity) and embodied (chip area × CPA, memory and storage per GB) carbon together; in a renewable-powered facility embodied carbon rises to 92-95%.

• Against published Google and Meta figures, error is 8.2% or less for training and -3.05% for embodied carbon.

SustainabilityEnergyPerformance metrics

Read / Download

All content and rights belong to the respective authors and publishers. SoSo CAD provides only the citation, its own summary and a link to the source, and is not responsible for the sources' content.