三菱重工业

LLM-as-a-judge systems are now routinely used for open-ended model evaluation, where human preference annotation is costly, slow, and difficult to reproduce. Yet these judges are often reported as scalar accuracy, win-rate, or agreement devices. We argue that a judge should instead be reported as a measurement instrument. We introduce a Judge Datasheet protocol that measures dark current under true-vacuum inputs, stable cross-sensitivity to same-quality surface variation, positional false preference, target sensitivity on a controlled quality ladder, and the criterion or operating point induced by tie instructions. The direction-stability decomposition reveals that apparent Delta0 preference can be stable surface response or disguised position bias. In a three-judge open-weight case study, Llama-3.1-8B shows high dark current and presentation-conflicted Delta0 behavior, Qwen2.5-14B is vacuum-clean and target-sensitive but mixes stable and positional over-discrimination, and Qwen2.5-32B is vacuum-clean with low stable cross-sensitivity and low positional false preference. A strict tie criterion eliminates Qwen32B Delta0 false preference but absorbs marginal Delta1 target signals into ties while preserving Delta5 sensitivity. The results show that prompting moves the criterion, not the resolution. We do not claim that the downstream mechanism hypothesis that motivated this work is confirmed; the contribution is a metrological protocol for measuring the measuring device before downstream claims are made.

Hydrogen is a promising zero-carbon fuel to achieve decarbonisation targets. Its broad flammability range compared to hydrocarbons allows fuel-leaner operations, which can help to mitigate thermal nitrogen oxides emissions. The Lean Direct Injection Combustor (LDIC) design is used to mitigate flame flashback arising under fully premixed conditions. For the LDIC used for this study, the hydrogen is injected transversely through two opposed jets into a cross-flowing air and the flame is stabilised using pilots. The global equivalence ratio of this main flame is varied by gradually changing the hydrogen mass flow rate until the lean blow-off (LBO) of the main flame occurs. The LES of these flames are conducted using flamelet based models for partially premixed combustion. The computed velocity statistics for non-reacting flow compare well with measurements. The computed flame attributes and LBO over a range of operating conditions are observed to agree well with the corresponding measurements. The physical insights obtained from the LES analyses are used to derive a correlation for the LBO, which agrees well with the measurements and LES results for two burner configurations.Novelty and significance statementThis study considers flame blow-off in a lean direct injection (LDI) burner using hydrogen, which is injected transversely into an air stream. The fuel is injected upstream of the combustor entry creating a partially premixed fuel-air mixture. The novelty of this work is the joint experimental and numerical investigation of stable flames and those close to blow-off limit. The second novelty is using LES insights to derive a blow-off correlation for the LDI burner since the classical correlations do not hold well. The role of friction velocity in the blow-off phenomenon of the LDI burner is demonstrated for the first time in this study.

Hydrogen/methane (H2/CH4) blends serve as a transitional fuel for gas turbines, offering a pathway toward cleaner energy by reducing carbon emissions while leveraging existing natural gas infrastructure. Understanding the effects of H2 content and pressure on the flame shape is essential for safe operation and optimising combustion performance. In this work, turbulent lean premixed Bunsen H2/CH4 flames are studied using large eddy simulation (LES), focusing on the flame brush length (fl) and thickness (fb) along the centreline. The results have been compared with available measurements for validation. The findings indicate that increasing H2 content leads to shorter and thinner flames, while the pressure effects are minimal. At the lower H2 content (≤30

Ammonia is a promising alternative to hydrocarbon fuels for achieving carbon neutrality. To suppress greenhouse gas emissions, carbon-free fuels including ammonia are expected to be implemented in practical combustors. In recent large-scale gas turbines, pressure ratios exceed 20; therefore, clarifying turbulent combustion characteristics under high-pressure conditions relevant to such system is necessary. In this study, the turbulent flames of ammonia/air mixtures stabilized on a nozzle burner are experimentally investigated in a high-pressure combustion chamber at pressures up to 2.0 MPa, and a mixture temperature of 573 K. Instantaneous flame images were acquired using OH-PLIF. Turbulent flame characteristics, including the local radius of curvature, which correlates with the representative flame wrinkling scale, and turbulent burning velocity, were evaluated under various pressure P and turbulence intensity u′ conditions. The results indicated that the flame front became finer with increasing pressure, and cusp-like structures appeared at high pressures. Flame-front characteristics were assessed based on the local radius of curvature. The variation in turbulent burning velocity ST, normalized by laminar burning velocity SL, namely ST/SL, was examined as a function of turbulence intensity u′. The observed changes in ST/SL were attributed to competition between variation in the Kolmogorov length scale and the characteristic length scales of flame instability. Finally, an empirical correlation for ST/SL was evaluated. The results showed that general correlations proposed in previous studies, such as ST/SL ∼ {(u′/SL)(P/P0)}0.38, did not apply to pure ammonia/air turbulent flames, particularly under high-pressure conditions. Accordingly, a modified empirical correlation, ST/SL ∼ (u′/SL)0.75(P/P0)0.23 was proposed. This correlation indicated that the effect of pressure on ST/SL was considerably weaker than that of u′/SL, as the power exponent of P/P0 is smaller than that of u′/SL.Novelty and significance statement: Turbulent flames stabilized on a nozzle burner were investigated for ammonia/air mixtures up to 2.0 MPa, relevant to large-scale gas turbine operation, using OH-PLIF measurements. To the best of the authors’ knowledge, this is the first turbulent-flame experiment conducted at pressures up to 2.0 MPa using a pure ammonia/air mixture. Flame images obtained by OH-PLIF were used to evaluate the characteristic length scales of the turbulent flame front and the turbulent burning velocity up to 1.5 MPa. These fundamental turbulent flame characteristics, particularly the turbulent burning velocity, are essential for the design of ammonia-fueled internal combustion systems, including gas turbines and reciprocating engines. These findings provide a basis for predicting the turbulent burning velocity of ammonia/air premixed flames under operational conditions relevant to large-scale gas turbine combustors, thereby supporting the design and optimization of ammonia-fueled internal combustion engines.

In a DC microgrid, the increasing presence of Constant Power Loads (CPL) may lead to instability in the DC voltage due to the CPL’s negative resistance characteristic. This paper presents an active voltage oscillation suppression method that utilizes an Oscillation Canceller (OC) for stabilization. The target system consists of four parallel CPLs powered by an AC/DC converter. A sequential verification framework—consisting of impedance-based stability analysis, Model-in-the-Loop (MIL) simulation, and Power Hardware-in-the-Loop (PHIL) simulation testing—is employed. Theoretical predictions are validated through MIL simulations, and the effectiveness of the OC is demonstrated. To address the challenge of accurately emulating CPL behavior in PHIL environments, a novel variant of the Damping Impedance Method (DIM), termed Admittance-based Partial Virtual DIM, is introduced. Experimental results confirm the feasibility and effectiveness of the proposed methodology.