Entropy Hunter: Fine-Tuning Qwen3-8B for Industrial Exergy Analysis

Large language models produce impressive results on general tasks. But can an 8-billion parameter model perform exergy analysis on industrial thermodynamic equipment? On a consumer GPU, with a startup budget?

Entropy Hunter was built to answer this question. It's a domain-specific LLM specialized for industrial equipment exergy analysis via LoRA fine-tuning on Qwen3-8B. At v0.4, it achieved a 92.7% benchmark score (Grade A-) at a $216 total cost.

What Is Exergy Analysis?

Exergy is the amount of "usable energy" a system possesses to the extent it is not in equilibrium with its surroundings. While energy is conserved, exergy is not — it is destroyed in every real process due to irreversibilities.

In industrial plants, exergy analysis reveals which equipment causes how much loss, how much of that loss is avoidable, and where improvement potential is concentrated. Traditional energy analysis answers "how much energy is there?" while exergy analysis answers "how much of that energy is actually usable?"

Version Evolution

The path from Base Qwen3-8B to v0.4 was not a straight line. Each version tested a different strategy, and the most critical lesson came from the v0.3 failure.

v0.1 (200 examples) applied initial fine-tuning — the model began learning exergy terminology but calculation consistency was low. v0.2 (600 examples) expanded the dataset, bringing the score to 78%.

v0.3 experimented by removing JSON output format — the model was forced to respond in free text. The result was dramatic: score dropped to 52.4%. This failure revealed that JSON format wasn't just output structure — it served as a scaffold for step-by-step reasoning. Each JSON field forced the model to complete a specific calculation step.

v0.4 applied the full pipeline with this insight: JSON scaffold preserved, quality control pipeline added, and trained on 1,235 validated examples. Result: 92.7%, Grade A-.

Training Pipeline

The pipeline consists of 5 stages:

1. Taxonomy: 7 equipment types, 48 subtypes, and 6 analysis families were systematically defined. This taxonomy ensured the training data comprehensively represented industrial scenarios.

2. Data Generation: 1,500 structured Q&A examples were generated via Claude Opus 4.6. Each example contained temperature, pressure, and flow rate values simulating real industrial conditions.

3. Quality Control: 8 thermodynamic validation checks were applied. Energy balance, mass conservation, entropy generation, and other physical constraints were verified for every example.

4. LoRA Fine-Tuning: Validated 1,235 examples were used to train LoRA adapters on Qwen3-8B. Rank 64, alpha 128, single consumer GPU (24GB VRAM) with 4-bit quantization.

5. Evaluation: Automated benchmark across 6 analysis families. Comparison against CoolProp reference values with ±2% tolerance.

Equipment & Analysis Coverage

Entropy Hunter covers 7 core industrial equipment types: turbine, compressor, pump, heat exchanger, boiler, mixing chamber, and nozzle/diffuser. These equipment types are broken into 48 subtypes — reflecting real industrial configurations like steam turbines, centrifugal compressors, and plate heat exchangers.

The 6 analysis families cover different aspects of exergy analysis:

• ED (Exergy Destruction): Irreversibility-driven exergy loss
• EE (Exergetic Efficiency): Component exergy utilization effectiveness
• AV (Avoidable): Exergy destruction share with improvement potential
• UN (Unavoidable): Minimum destruction within technological limits
• EI (Improvement Potential): Van Gool improvement potential calculation
• EGM (Entropy Generation Minimization): Optimization via Bejan's EGM method

Benchmark Results

v0.4 performance across the 6 analysis families:

100% score on AV and UN families — showing the model fully learned avoidable/unavoidable exergy destruction decomposition. These families offer clear formula frameworks that, combined with JSON scaffold, produce consistent results.

ED (96.2%) and EE (94.5%) — exergy destruction calculation and efficiency evaluation performed with high accuracy. These categories require direct application of standard thermodynamic formulas.

EI (93.5%) — Van Gool improvement potential, a metric combining efficiency and exergy destruction values. The model handles this multi-step calculation well.

EGM (72.0%) — the weakest area. Entropy generation minimization requires abstract optimization reasoning, unlike other categories. The model is consistent here but accuracy is lower — likely at the capacity limits of an 8B model.

Compared to Base Qwen3-8B, an average +54.7 point improvement; compared to v0.2, a +14.7 point increase.

Quality Control

Data quality is the foundation of model quality. Of 1,500 generated examples, 1,235 passed all 8 thermodynamic validation checks (82.3% validity rate).

The 8 checks:

1. Energy Balance (ΔE = Q − W) — first law consistency
2. Mass Conservation (Σṁᵢₙ = Σṁₒᵤₜ) — inlet-outlet mass flow balance
3. Entropy Generation (Ṡgen ≥ 0) — second law violation check
4. Exergy Balance (Ėd ≥ 0) — negative exergy destruction check
5. Temperature Range (T > 0 K) — physically meaningful temperatures
6. Pressure Positivity (P > 0) — physically meaningful pressures
7. Efficiency Bounds (0 ≤ η ≤ 1) — physically possible efficiency values
8. Second Law Compliance (COP ≤ COPCarnot) — Carnot limit not exceeded

The 265 failed examples (17.7%) typically failed on energy balance and entropy generation checks — showing that even Opus 4.6 occasionally produces inconsistent results in complex multi-step thermodynamic calculations.

Key Findings

Known Limitations

Known limitations of Entropy Hunter v0.4:

• Numerical JSON output dependency: The model performs poorly without structured JSON. Cannot perform exergy analysis in free text format — this is an 8B model limitation.
• Arithmetic variance: The same question can produce ±2-3% different numerical results across runs. Use T=0.0 for deterministic results, but this reduces diversity.
• EGM weakness: Entropy generation minimization at 72% is the weakest area. Abstract optimization reasoning is at the limits of 8B model capacity.
• Steam table approximation: The model uses approximate values learned from training data rather than precise databases like CoolProp or REFPROP. Accurate within ±2% tolerance, but CoolProp reference should be preferred for engineering design.

Methodology

• Base model: Qwen3-8B (Qwen/Qwen3-8B)
• Fine-tuning method: LoRA (Low-Rank Adaptation)
• LoRA configuration: rank=64, alpha=128, dropout=0.05
• Quantization: 4-bit (QLoRA, bitsandbytes)
• Hardware: Single consumer GPU, 24GB VRAM
• Training time: ~4 hours
• Training data: 1,235 validated examples (from 1,500 generated)
• Data generation: Claude Opus 4.6 API
• Reference library: CoolProp 7.2.0
• Tolerance: ±2% (industrial engineering standard)
• Total cost: $216 (data generation ~$180, fine-tuning ~$28, evaluation ~$8)
• Thinking mode: Disabled (better performance on fine-tuned version)
• Temperature: 0.7 (optimal balance)
• Benchmark: 6 analysis families (ED, EE, AV, UN, EI, EGM)

Resources

• Model: HuggingFace — olivenet/entropy-hunter-v0.4
• Source code: GitHub — olivenet-iot/entropy-hunter
• CoolProp: coolprop.org
• Qwen3-8B: HuggingFace — Qwen/Qwen3-8B
• LoRA: HuggingFace PEFT