排出特性評価とユースケース応用のための物理情報デジタルフレームワーク

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:PINNと大気逆解析を統合した排出量推定の実証ベンチマークとして、測定ベースMRV研究の設計参考になる。

🏢実務担当者:自社施設のScope1排出を連続監視・検証するデジタルツイン導入の実現性と精度水準を判断する材料になる。

🏛政策担当者:測定ベース排出報告の制度設計において、センサー＋物理モデル手法の精度と限界（CO2限定）を把握する参考になる。

Emissions quantification and production surveillance are commonly implemented as separate measurement workflows, relying on dedicated gas sensors, intrusive instrumentation, or periodic surveys that are difficult to scale across heterogeneous facilities and changing operating conditions. This paper presents a physics-informed emissions-quantification framework that couples upstream fuel metering with remote atmospheric inversion to reconstruct facility-level carbon flux continuously. The present study validates the framework for carbon dioxide (CO2); methane-specific detection and quantification performance are not evaluated. The framework is implemented as a digital twin that embeds mass conservation, combustion stoichiometry, and advection-diffusion transport within a physics-informed neural network (PINN) architecture. Facility emissions are represented as latent source terms in a constrained mass-balance formulation. Emission rates are inferred from residual structure between (i) measured fuel input and thermodynamically consistent inventory/outlet states and (ii) atmosphere-derived concentration fields. The formulation accounts for measurement sparsity, sensor noise, and transient atmospheric variability while enforcing physical consistency and reporting observability diagnostics. Field testing was conducted during 2025 at an operating industrial manufacturing facility under building-scale flow obstruction, variable wind, and complex plume transport. A 12-station distributed sensing array spanning approximately 11 m end to end across a rooftop transect measured CO2 and volatile organic compound (VOC) concentration gradients during controlled-combustion and controlled-release scenarios. Upstream fuel input was quantified using metering augmented by PINN-based estimation to provide the process-side carbon reference used in the coupled estimator. Two evaluations were kept distinct: controlled CO2 releases, for which the meter-side constraint was omitted and release metadata were withheld, provided an independent end-to-end check of the sensing-and-inversion chain; the full-year atmosphere-versus-fuel comparison assessed closed-loop consistency because the fuel-derived reference was included as a soft training constraint. Twenty controlled CO2 releases from calibrated cylinders, conducted at 0.5–40 kg/h from three source locations within a 10 m radius, yielded a maximum absolute relative error of 10.3% (mean 5.5%), a root-mean-square error (RMSE) of approximately 1.0 kg/h, and R2 = 0.992. In the 2025 closed-loop comparison, monthly signed differences for months with a nonzero fuel reference were bounded within ±14.5%, while the annual aggregate atmosphere-versus-fuel difference was +6.8%. The controlled-release and supplementary characterization set covered 0–45 km/h; the broader full-year field record spanned 3–58 km/h, and the 0–1 km/h calm bin was laboratory-only. Descriptive wind-bin means showed the smallest absolute bias (≤4%) from 12 to 38 km/h. Because within-bin sample counts and dispersion were unavailable for this analysis, that range is presented as a provisional low-bias regime rather than a statistically validated operating envelope. All reported accuracy results are CO2-specific. Methane deployment requires methane-sensitive instrumentation, controlled methane releases, and site-specific qualification.

• semanticscholar https://doi.org/10.2118/234192-msfirst seen 2026-10-17 05:08:12 · last seen 2026-11-06 05:16:22