Paper reviews organized by application systems and algorithmic methods. Just as elite athletes train every day, serious researchers read strong papers every day.응용 시스템과 알고리즘 방법론으로 나누어 정리한 논문 리뷰입니다. 좋은 연구자가 되기 위한 매일의 연구 훈련으로, 핵심 주장과 한계를 함께 읽습니다.
Application Reviews응용 시스템 리뷰
Reviews grouped by the systems where optimization, AI, and decision tools are applied.최적화, AI, 의사결정 도구가 적용되는 시스템별로 묶은 리뷰입니다.
Reviews grouped by the computational method, with emphasis on assumptions, evidence, and feasibility-critical limits.계산 방법론별로 묶은 리뷰이며, 가정, 근거, 실행 가능성의 한계를 함께 봅니다.
With 30 samples in hand, NOMAD can fit a local model to propose the next evaluation. When does one Search call save work in Poll?블랙박스 함수의 평가값 30개로 다음 점을 고르는 과정을 따라가며, Search에 쓴 평가 한 번이 어떤 조건에서 Poll 비용을 줄이는지 계산합니다.
Algorithm 1027: NOMAD Version 4: Nonlinear Optimization with the MADS AlgorithmAlgorithm 1027: NOMAD Version 4: Nonlinear Optimization with the MADS Algorithm
Separate the prediction horizon from the integer horizon: retain a long look ahead while relaxing distant on/off decisions to reduce computation.예측기간은 길게 유지하면서 가까운 미래의 on/off만 정수로 결정하고, 먼 미래에는 연속 완화를 허용해 계산 부담을 줄이는 아이디어.
Real-Time Nonlinear Mixed-Integer Model Predictive Control With Horizon-Tail RelaxationsReal-Time Nonlinear Mixed-Integer Model Predictive Control With Horizon-Tail Relaxations
What a conditional miss bound means, how certification costs grow, and how sampled scenario cuts support mixed-integer robust optimization.조건부 놓침 확률의 의미와 인증 검사량을 계산하고, 오류 예산 배분과 정수 설계별 Markov chain을 Benders cut에 연결합니다.
Soft Separation for Adaptive Robust OptimizationSoft Separation for Adaptive Robust Optimization
How scenario visit frequencies determine the optimization direction, why one-step Markov samples are biased, and which cumulative-error bound yields convergence.시나리오 방문 비율이 최적화 방향을 결정하는 이유를 살펴보고, 한 번의 Markov 전이로 생기는 편향과 수렴에 필요한 누적 오차 상계를 유도합니다.
A constraint can block all feasible motion while its gradient says nothing. Ramp limits and short multiplier proofs separate KKT necessity, convex sufficiency, and optimizer differentiability.제약은 실제 움직임을 막는데 gradient는 이를 표현하지 못할 수 있다. Ramp 예제와 짧은 승수 증명으로 KKT의 필요성, convex 문제의 충분성, 최적해의 미분 가능성을 구분한다.
A hard ramp example separates function-value gradients from action sensitivity, then derives forward and backward passes through the KKT equations of a quadratic program.Hard ramp 예제로 함수값의 미분과 action 민감도를 구분하고, QP의 KKT 조건을 통해 forward·backward pass를 유도한다.
OptNet: Differentiable Optimization as a Layer in Neural NetworksOptNet: Differentiable Optimization as a Layer in Neural Networks
A constrained optimizer cannot react to every prediction error. PEAR keeps the directions that can move the decision, making it most natural when system dynamics and constraints stay fixed while objective coefficients change across instances.제약 최적화기는 모든 예측 오차에 반응하지 않는다. PEAR는 의사결정을 움직일 수 있는 오차만 남기며, 시스템의 동역학과 제약은 같고 목적함수 계수만 달라지는 문제에 특히 잘 맞는다.
Decision-Focused Learning via Tangent-Space Projection of Prediction ErrorDecision-Focused Learning via Tangent-Space Projection of Prediction Error
CAffNet builds input-dependent affine constraints into a neural output layer. It guarantees constraint satisfaction when the supplied feasible set is nonempty. Enumerating active constraint combinations can be costly, and physical safety still depends on the accuracy of the constraint model.CAffNet은 입력에 따라 달라지는 아핀 제약을 신경망 출력층에 넣는다. 주어진 제약을 만족하는 출력이 존재하면 제약 만족을 보장한다. 다만 활성 제약 조합을 열거하는 비용이 크고, 물리적 안전은 제약 모델의 정확성에 달려 있다.
CAffNet: Hard Constraint-Affine Neural NetworksCAffNet: Hard Constraint-Affine Neural Networks
How online MDP estimation, interval uncertainty, and probabilistic shielding form a coupled loop in which blocking risky actions can also block the evidence needed to learn safety.온라인 MDP 추정, 구간 불확실성, 확률적 실딩이 결합될 때 위험한 행동의 차단이 안전을 학습하는 데 필요한 증거까지 막는 교착을 분석한다.
A critical note on learning interpretable objective weights from expert production plans while preserving a known industrial MILP.알려진 산업 MILP를 유지하면서 전문가 생산계획으로부터 해석 가능한 목적함수 가중치를 학습하는 inverse optimization에 대한 비판적 정리.
Uncovering expert objectives in production planning via inverse optimization: An industrial case studyUncovering expert objectives in production planning via inverse optimization: An industrial case study
Diffusion models need the marginal score to reverse a noising process. This note explains why a tractable conditional Gaussian target learns that score and in what sense the two losses are equivalent.Diffusion model이 noising process를 역전할 때 필요한 marginal score를 조건부 Gaussian target으로 학습할 수 있는 이유와 두 손실함수의 정확한 등가성을 정리한다.
Stanford CME296 Diffusion & Large Vision Models | Spring 2026 | Lecture 2 - Score matchingStanford CME296 Diffusion & Large Vision Models | Spring 2026 | Lecture 2 - Score matching
A solvent with the lowest nominal reboiler duty may still require the tightest circulation control. This note separates energy efficiency, disturbance sensitivity, and ease of operation in an Aspen Plus comparison of MEA and MDEA/PZ solvents.정상 설계점에서 regeneration energy가 가장 낮은 흡수제가 반드시 운전하기 쉬운 것은 아니다. Aspen Plus 기반 MEA 및 MDEA/PZ 비교를 통해 에너지 효율, 외생조건 민감도, 운전 오차 민감도를 구분한다.
Robust process design and operation for efficient CO₂ capture under variable supply-and-demand conditionsRobust process design and operation for efficient CO₂ capture under variable supply-and-demand conditions
The paper's strongest contribution is not attaching a neural network to a distillation model. It is the combination of precise error localization, a differentiable industrial-scale column solver, and a deliberately small thermodynamic correction network.이 논문의 가장 강한 기여는 증류 모델에 neural network를 붙였다는 사실이 아니다. 정확한 error localization, 미분 가능한 산업 규모 column solver, 최소한의 thermodynamic correction을 하나의 구조로 묶었다는 데 있다.
Hybrid modeling of an industrial LPG debutanizer using a differentiable first-principles distillation solver with real plant dataHybrid modeling of an industrial LPG debutanizer using a differentiable first-principles distillation solver with real plant data
A critical note on ChatP&ID: P&IDs are better treated as structured engineering knowledge graphs than as raw images or raw XML, but the benchmark mainly validates context engineering rather than a new GraphRAG algorithm.ChatP&ID에 대한 비판적 노트. P&ID는 이미지나 raw XML보다 구조화된 engineering knowledge graph로 다루는 편이 설득력 있지만, 이 논문의 핵심은 새로운 GraphRAG algorithm보다 context engineering과 retrieval 구조 설계에 가깝다.
GraphRAG for Engineering Diagrams: ChatP&ID Enables LLM Interaction with P&IDsGraphRAG for Engineering Diagrams: ChatP&ID Enables LLM Interaction with P&IDs
A note on Zhao and Fink's H-CDE model for separating slow latent degradation from fast operational dynamics, read from the viewpoint of chemical plant monitoring and operation.Zhao와 Fink의 H-CDE 모델을 chemical plant monitoring과 operation 관점에서 읽는다. 핵심은 느린 latent degradation과 빠른 operational dynamics를 분리해 추론하는 구조다.
Disentangling slow and fast temporal dynamics in degradation inference with hierarchical differential modelsDisentangling slow and fast temporal dynamics in degradation inference with hierarchical differential models
A critical note on Tolerance Ball acquisition: a clean probability-of-feasibility objective for specification-driven inverse design, but not a direct optimizer of diversity, boundary coverage, or global feasible-set recovery.Specification-driven inverse design에서 Tolerance Ball acquisition을 비판적으로 읽는다. TB는 valid hit probability에는 잘 정렬되어 있지만 diversity, boundary coverage, global feasible-set recovery를 직접 최적화하지는 않는다.
Range-aware Bayesian optimization for discovering diverse designs within target property windowsRange-aware Bayesian optimization for discovering diverse designs within target property windows
A note on Cortés-Peña and Zavala's phenomena-based graph representation for flowsheet simulation, where nonlinear thermodynamic blocks are separated from process-wide linear material and energy balance solves.Cortés-Peña와 Zavala의 현상 기반 그래프 표현을 정리한다. 핵심은 비선형 열역학 블록과 전 공정 수준의 선형 물질수지 및 에너지수지 풀이를 분리하는 것이다.
Phenomena-based graph representations and applications to chemical process simulationPhenomena-based graph representations and applications to chemical process simulation
A critical note on a BOHB-MILP framework for international liquid hydrogen supply-chain design under hourly renewable variability, weekly shipping, lead time, and sampled demand-weather scenarios.시간별 재생에너지 변동, 주별 선박운항, 운송 lead time, 수요-기상 시나리오를 결합한 국제 액화수소 공급망 BOHB-MILP 프레임워크에 대한 비판적 노트.
Techno-economic analysis for design and management of international green hydrogen supply chain under uncertainty: An integrated temporal planning approachTechno-economic analysis for design and management of international green hydrogen supply chain under uncertainty: An integrated temporal planning approach
A note on a Max-Cut solver that replaces repeated SDP relaxation solves with a feasibility-preserving GNN surrogate, keeping branch-and-bound correctness by constructing dual-feasible upper bounds.Max-Cut branch-and-bound에서 반복되는 SDP relaxation solve를 feasibility-preserving GNN surrogate로 대체하되, dual-feasible upper bound를 만들어 exactness를 유지하는 접근을 정리한다.
A critical note on an off-grid PV-battery-PEM hydrogen TEA that folds PEM degradation, battery fade, replacement timing, and Sobol/XGBoost/SHAP interpretation into the sizing objective.PV-battery-PEM 오프그리드 그린수소 시스템에서 PEM degradation, battery fade, replacement timing, Sobol/XGBoost/SHAP 해석을 sizing objective에 함께 넣은 TEA 논문에 대한 비판적 정리.
Explainable degradation-aware techno-economic optimization of off-grid green hydrogen productionExplainable degradation-aware techno-economic optimization of off-grid green hydrogen production
A critical note on Q/K/V projection sharing in Transformer attention, where K=V preserves query-key directionality while cutting KV-cache memory in half.Transformer attention에서 Q/K/V projection을 모두 독립적으로 둘 필요가 있는지 검토하고, K=V sharing이 query-key 방향성을 유지하면서 KV cache를 절반으로 줄이는 압축 축임을 비판적으로 정리한다.
Do Transformers Need Three Projections? Systematic Study of QKV VariantsDo Transformers Need Three Projections? Systematic Study of QKV Variants
A critical note on TalkToAgent, an LLM multi-agent framework that reuses existing LLMs to route natural-language questions to XRL tools for chemical process-control policies.기존 LLM을 화학 도메인에 맞게 fine-tuning하는 대신, 자연어 질문을 XRL 도구와 simulator에 연결해 chemical process-control RL policy를 설명하는 TalkToAgent에 대한 비판적 정리.
TalkToAgent: A multi-agent LLM Framework for natural language explanation of reinforcement learning policiesTalkToAgent: A multi-agent LLM Framework for natural language explanation of reinforcement learning policies
A note on ORACLE, which turns near-optimal energy-system exploration from point generation into an inner/outer approximation problem with a certified distance metric.near-optimal energy-system exploration을 단순한 point generation이 아니라 inner/outer approximation과 certified distance metric의 문제로 바꾸는 ORACLE에 대한 정리.
ORACLE: A rigorous metric and method to explore all near-optimal designs for energy systemsORACLE: A rigorous metric and method to explore all near-optimal designs for energy systems
A critical note on how fixed-efficiency electrolyzer models and low-resolution renewable data can bias green-hydrogen LCOH estimates.고정 효율 전해조 모델과 낮은 해상도의 재생에너지 데이터가 그린수소 LCOH 추정에 어떤 bias를 만드는지 비판적으로 정리한다.
Impact of electrolyzer-model fidelity and renewable-data resolution on techno-economic assessments of green hydrogen systemsImpact of electrolyzer-model fidelity and renewable-data resolution on techno-economic assessments of green hydrogen systems
A critical note on why renewable green-hydrogen TEA should treat electrolyzer degradation, replacement, and on/off operation as endogenous economic effects rather than fixed lifetime parameters.재생에너지 기반 그린수소 경제성 분석에서 전해조 degradation, 교체주기, on/off 운전을 고정 수명 파라미터가 아니라 경제성을 바꾸는 내생적 효과로 보아야 하는 이유를 정리한다.
The impact of degradation on the economics of green hydrogenThe impact of degradation on the economics of green hydrogen
A critical note on modeling data centers as restless bandit arms for grid demand response, and on where the finite-state RMAB abstraction is useful or fragile.데이터센터를 전력망 demand response를 위한 restless bandit arm으로 모델링하는 접근을 정리하고, finite-state RMAB 추상화가 유용한 지점과 취약한 지점을 비판적으로 검토한다.
Robust Restless Multi-Armed Bandit for Data Center Flexibility Services Through Virtual Machine SchedulingRobust Restless Multi-Armed Bandit for Data Center Flexibility Services Through Virtual Machine Scheduling
A critical note on MadNCL, which combines Algorithm NCL, MadNLP, and GPU-friendly KKT reformulations to improve robustness on large-scale degenerate nonlinear programs.Algorithm NCL, MadNLP, GPU 친화적 KKT 재구성을 결합해 대규모 퇴화 비선형계획에서 강건성을 높이려는 MadNCL 논문에 대한 비판적 정리.
MADNCL: a GPU implementation of algorithm NCL for large-scale, degenerate nonlinear programsMADNCL: a GPU implementation of algorithm NCL for large-scale, degenerate nonlinear programs
A critical note on NLPOpt-Net, which learns a parametric NLP solution map with a neural warm start and an objective-aware differentiable projection layer.NLPOpt-Net을 신경망 warm start와 목적함수 인식 differentiable projection layer가 결합된 파라메트릭 NLP solution-map 학습법으로 정리한 비판적 노트.
NLPOpt-Net: A Learning Method for Nonlinear Optimization with Feasibility GuaranteesNLPOpt-Net: A Learning Method for Nonlinear Optimization with Feasibility Guarantees
LoRA reduces fine-tuning cost by learning low-rank task-specific corrections on top of frozen pretrained weights, while QLoRA adds 4-bit quantization so larger base models can be adapted under tighter GPU memory constraints.LoRA는 고정된 pretrained weight 위에 저랭크 task-specific correction만 학습해 fine-tuning 비용을 줄이고, QLoRA는 여기에 4-bit quantization을 결합해 더 큰 base model을 제한된 GPU memory에서 적응시킬 수 있게 한다.
A critical note on enforcing input-dependent linear equality and inequality constraints in neural network outputs using a robustly feasible decision-rule anchor and minimal interpolation.강건하게 실행가능한 결정규칙 앵커와 최소 보간을 이용해 신경망 출력의 입력 의존 선형 등식 및 부등식 제약을 강제하는 방법에 대한 비판적 연구 노트.
Enforcing hard linear constraints in deep learning models with decision rulesEnforcing hard linear constraints in deep learning models with decision rules
FlashAttention is not an approximation to attention. Its core idea is to avoid materializing the N by N attention matrix in HBM by computing tiled attention in SRAM and maintaining online softmax statistics.FlashAttention은 어텐션의 근사가 아니다. 핵심은 SRAM에서 타일 단위 어텐션을 계산하고 online softmax 통계를 유지함으로써 HBM에 N by N 어텐션 행렬을 만들지 않는 것이다.
A critical note on learned Lyapunov terminal costs for NMPC, focusing on Cholesky-structured positive-definite surrogates, horizon compression, and the unresolved role of approximation error.NMPC의 학습 기반 Lyapunov terminal cost를 비판적으로 읽는 글로, Cholesky 구조의 양의 정부호 surrogate, horizon compression, 그리고 approximation error의 미해결 역할을 중심으로 다룬다.
Learning Lyapunov terminal costs from data for complexity reduction in nonlinear model predictive controlLearning Lyapunov terminal costs from data for complexity reduction in nonlinear model predictive control
This note reads Lyapunov-based safe policy optimization as a practical projection bridge from finite CMDP safe policy iteration to continuous-action deep reinforcement learning, while separating exact CMDP guarantees from local approximation behavior.이 노트는 Lyapunov 기반 안전 정책 최적화를 유한 CMDP의 안전 정책 반복과 연속 행동 심층 강화학습을 잇는 실용적 투영 구조로 읽는다. 동시에 정확한 CMDP 보장과 국소 근사에서의 경험적 안전성을 구분한다.
Lyapunov-based Safe Policy Optimization for Continuous Control연속 제어를 위한 Lyapunov 기반 안전 정책 최적화
Lyapunov constraints turn an expected cumulative safety budget in a CMDP into local restrictions on policy improvement. This note examines the exact certificate logic and the weaker status of neural approximations.Lyapunov 제약은 CMDP의 기대 누적 안전 예산을 정책 개선의 국소 제약으로 바꾼다. 이 노트는 정확한 인증 논리와 신경망 근사에서 약해지는 보장 수준을 구분해 읽는다.
A lyapunov-based approach to safe reinforcement learning안전 강화학습을 위한 Lyapunov 기반 접근