💬 NLP

Full-Stack Domain Enhancement for Combustion LLMs: Construction and Optimization

This paper introduces the first full-stack domain-enhanced workflow for combustion science that integrates automated corpus construction, incremental pre-training, instruction fine-tuning, and verifiable reward-based reinforcement learning to eliminate hallucinations and ensure adherence to physical laws, validated by the new FlameBench benchmark which demonstrates superior performance over state-of-the-art models.

Original authors: Quanjia Xiao, Weimin Ouyang, Zonglin Yang, Tianhao Wu, Qingguo Zhou, Runze Mao, Zhi X. Chen

Published 2026-03-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Quanjia Xiao, Weimin Ouyang, Zonglin Yang, Tianhao Wu, Qingguo Zhou, Runze Mao, Zhi X. Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant, well-read student named Alex. Alex has read millions of books, knows how to write essays, and can chat about almost anything. However, if you ask Alex to design a rocket engine or explain why a specific fuel mixture might explode, Alex starts making things up. Alex might say, "Well, fire needs water to burn," or invent a law of physics that doesn't exist. This is called hallucination.

The problem is that Alex was trained on general books (history, news, fiction) and doesn't truly understand the strict, unbreakable rules of combustion science (how fire, fuel, and air interact).

This paper introduces a new training program to turn Alex into a Master Combustion Engineer. Here is how they did it, explained simply:

1. The Problem: The "Generalist" Trap

General AI models are like general practitioners. They are great at chatting, but when you ask them about complex engineering, they guess. In combustion, guessing is dangerous. If you violate the laws of physics (like conservation of energy), your prediction is useless. Existing methods tried to fix this by giving the AI a "textbook" to look at while answering (called RAG or Retrieval-Augmented Generation), but this is like giving a student a dictionary during a math test—they still might not understand how to solve the problem, they just copy the words.

2. The Solution: A "Full-Stack" Training Camp

The authors built a complete, four-step training camp to rewire the AI's brain specifically for fire and fuel.

Step 1: The Specialized Library (Corpus Construction)

Instead of just reading random books, they built a massive, specialized library containing hundreds of thousands of scientific papers, textbooks, and code specifically about combustion.

Analogy: Imagine taking a student who only reads comic books and forcing them to read every engineering manual, chemistry textbook, and physics paper ever written about fire. They also mixed in some general reading so the student doesn't forget how to speak normal English.

Step 2: The "Deep Dive" (Incremental Pre-Training)

The AI was made to "read" this new library from start to finish. This isn't just skimming; it's absorbing the vocabulary, the formulas, and the core concepts.

Analogy: This is like the student moving into a dorm room filled with engineering diagrams. They start dreaming in equations and stop thinking of fire as just "hot stuff" and start seeing it as "chemical reactions and fluid dynamics."

Step 3: The "Drill Sergeant" (Supervised Fine-Tuning)

Now that the AI knows the facts, it needs to learn how to think. The researchers gave the AI thousands of examples of experts solving problems step-by-step (Chain of Thought).

Analogy: The student is now in a classroom where a strict professor shows them exactly how to solve a problem: "First, check the fuel ratio. Second, calculate the heat. Third, check the pressure." The student learns the process, not just the answer.

Step 4: The "Physics Police" (Reinforcement Learning with Verifiable Rewards)

This is the most important part. The AI was put in a simulation where it had to solve complex problems. If it gave an answer that broke the laws of physics (e.g., creating energy out of nothing), it got a "punishment" (negative reward). If it followed the rules, it got a "praise" (positive reward).

Analogy: Imagine a driving simulator. If the student drives the car through a wall (violating physics), the game crashes and they get a "Game Over." If they drive correctly, they get points. Over time, the student learns that breaking the rules is impossible, so they stop trying to guess and start calculating correctly.

3. The Result: The "FlameBench" Test

To prove it worked, the authors created a special exam called FlameBench. It's a test full of tricky, real-world combustion questions that require deep reasoning.

The Old Way (General AI): Got about 27% right.
The "Textbook Lookup" Way (RAG): Got about 26% right (surprisingly, looking up answers didn't help much because the questions required connecting many dots).
The New "Full-Stack" AI: Got 43.8% right.

While 43% might sound low, in the world of complex scientific reasoning, this is a massive leap. More importantly, the new AI didn't just memorize answers; it learned to think like a scientist. It stopped making up fake physics laws.

Why This Matters

This paper shows that to make AI useful for hard sciences (like building rockets, designing new medicines, or optimizing energy), you can't just give it a search engine. You have to retrain its brain to internalize the rules of the universe.

The Big Takeaway:
They built a system that turns a "chatty" AI into a "reliable engineer" by feeding it specialized knowledge, teaching it how to reason step-by-step, and punishing it whenever it breaks the laws of physics. This paves the way for AI assistants that can actually help scientists discover new things without making dangerous mistakes.

1. Problem Statement

General-purpose Large Language Models (LLMs) face significant limitations when applied to combustion science, a field characterized by complex physical systems, strict conservation laws, and multi-physics interactions. The primary challenges identified are:

Domain Knowledge Deficit: General training corpora lack sufficient coverage of specialized concepts, equations, and modeling paradigms (e.g., reaction mechanisms, turbulent combustion models), leading to poor interpretation of symbolic systems.
Physical Inconsistency: Without explicit awareness of physical laws (mass, momentum, energy conservation), general LLMs generate "hallucinations" that are textually plausible but physically impossible (e.g., violating chemical kinetics or energy conservation).
Limitations of Existing Solutions:
- Retrieval-Augmented Generation (RAG): While helpful for knowledge sparsity, RAG fails to ensure consistency during multi-step reasoning and struggles with integrating heterogeneous knowledge fragments.
- Single-Stage Adaptation: Traditional domain adaptation often focuses on a single training stage (e.g., only fine-tuning), which is insufficient for robust deployment in engineering sciences requiring coupled dynamical reasoning.

2. Methodology

The authors propose a full-stack domain-enhanced workflow tailored specifically for combustion science. This framework integrates data construction, multi-stage model optimization, and standardized evaluation.

A. Corpus Construction

Scale: A mixed corpus of approximately 30 billion tokens.
Composition:
- 5B tokens: Combustion-specific data (English/Chinese academic publications, encyclopedic resources on physics/chemistry/engineering).
- 25B tokens: General pre-training data (to retain general language understanding and commonsense reasoning).
Processing: An automated pipeline handles PDF parsing, multi-level deduplication (exact and approximate via MinHash), and quality control (perplexity and relevance scoring) to produce high-quality structured data.

B. Multi-Stage Model Adaptation

The workflow proceeds through three distinct optimization stages:

Continued Pre-Training (CPT): The base model is trained on the mixed corpus for one epoch using a conservative learning rate. This injects domain terminology and fundamental physical relationships while preventing catastrophic forgetting of general capabilities.
Supervised Fine-Tuning (SFT):
- Phase 1: Alignment with general instruction-following behavior (800K samples).
- Phase 2: Adaptation to combustion-specific tasks using 12K Chain-of-Thought (CoT) examples covering knowledge queries, formula derivation, and experimental analysis.
Reinforcement Learning with Verifiable Rewards (RLVR):
- Based on the GRPO framework.
- Uses a binary reward function to explicitly penalize violations of domain knowledge and physical consistency.
- Includes KL divergence regularization to maintain training stability.
- Trained on 7K complex reasoning instances targeting multi-parameter coupling.

C. Evaluation Benchmark: FlameBench

A standardized benchmark specifically designed for combustion science reasoning.
Source: Extracted from peer-reviewed literature, dissertations, and code repositories.
Content: 436 high-quality questions covering eight subfields, each grounded in a unique source reference for reproducibility.
Validation: Questions undergo a three-stage validation process (difficulty filtering, correctness verification, expert calibration).

3. Key Contributions

Large-Scale Combustion Corpus: Construction of a high-quality, 30B-token dataset specifically derived from scientific publications, addressing the scarcity of domain-specific training data.
Full-Stack Workflow: Development of an integrated pipeline (CPT $\rightarrow$ SFT $\rightarrow$ RLVR) that simultaneously addresses knowledge gaps and enforces physical consistency, moving beyond single-stage adaptation.
FlameBench: The release of the first standardized evaluation benchmark for complex reasoning tasks in combustion science, enabling rigorous comparison of domain-specific models.
RLVR for Physical Laws: Demonstration that Reinforcement Learning with Verifiable Rewards can effectively stabilize policy optimization and reduce physically inconsistent generations in scientific tasks.

4. Experimental Results

The proposed model (optimized via CPT-SFT-RLVR) was evaluated against the base model, various training stages, and state-of-the-art baselines (including closed-source models and RAG systems) on FlameBench.

Performance Progression:
- Base Model (Qwen-8B): 26.8% accuracy.
- CPT: Improved to 33.3% (+6.5%), proving the efficacy of hybrid-corpus pre-training.
- SFT-General: Marginal gain (33.5%).
- SFT-Combustion: Further improvement to 35.1%, showing the value of domain-specific CoT data.
- RLVR-Opt: Achieved the highest accuracy of 43.8%, a significant jump from the SFT baseline.
Comparison with RAG:
- The RLVR-Opt model outperformed the best RAG baseline (RAG + GLM-4, 32.09%) by 11.71 percentage points.
- It surpassed the average RAG performance (26.24%) by a wide margin, demonstrating that internalized knowledge is superior to external retrieval for multi-step physical reasoning.
Comparison with General LLMs:
- The optimized model outperformed top-tier closed-source models like GPT-5 (15.6%), Gemini Pro (32.10%), and DeepSeek-R1 (28.37%).
Training Stability:
- Unlike vanilla RL which suffered from "shortcut-driven policy collapse" (response length dropping from ~~1,500 to ~100 tokens), the RLVR model initialized with SFT-Combustion maintained a stable response length (~~2,000 tokens) and higher validation accuracy, indicating convergence to a more deterministic and physically grounded reasoning process.

5. Significance and Impact

Scientific Reliability: This work establishes a technical foundation for developing scientific research agents that can reliably reason under explicit physical constraints, a critical requirement for engineering applications.
Paradigm Shift: It demonstrates that for hard-science domains, internalizing domain knowledge and physical laws through full-stack training is more effective than relying on retrieval mechanisms or general-purpose models.
Generalizability: The proposed workflow offers a generalizable paradigm for other engineering and scientific fields characterized by strict physical laws and multi-disciplinary coupling (e.g., fluid dynamics, materials science).
Resource Availability: The release of the FlameBench benchmark and the methodology for corpus construction provides essential resources for the broader AI4S community to advance domain-specific LLM development.