AI Research

AI Research Projects

Measuring and advancing AI performance in industrial thermodynamics through open-source benchmarks and domain-specific models.

Entropy Hunter

Domain-Specific LLM for Exergy Analysis

A domain-specific LLM specialized in exergy analysis of industrial equipment via LoRA fine-tuning on Qwen3-8B. Achieved 92.7% benchmark score (Grade A-) with only $216 total cost on a single consumer GPU.

92.7%
Benchmark Score
$216
Total Cost
48
Equipment Subtypes
4h
Training Time

Key Findings

  • 1JSON format acts as a reasoning scaffold — removing it drops score from 92.7% to 52.4%
  • 2Frontier-model-level domain performance on a consumer GPU for $216
  • 3Thinking mode disabled yields better results — conflicts with JSON scaffold
  • 4EGM at 72% is the weakest area — abstract optimization reasoning is the limit of 8B

ThermoQA

Thermodynamic AI Benchmark

An open-source, multi-tier benchmark system measuring AI models' thermodynamic problem-solving competency. Tests property lookups, component analysis, and cycle analysis.

Tier 1: Property Lookups

110 questions, single fluid (water), steam table value reading with CoolProp 7.2.0 reference.

Gemini 97.3% > GPT 96.9% > Opus 95.6% > DeepSeek 89.5%

Tier 2: Component Analysis

101 questions, 7 components, 3 fluids. Multi-step thermodynamic reasoning including energy balance, isentropic efficiency, and exergy.

Opus 92.0% > GPT 91.0% > Gemini 89.5% > DeepSeek 86.9%

Tier 3: Cycle Analysis

82 questions, 9 cycle types, 4 fluids. Full Rankine, Brayton, VCR, and CCGT cycle analysis with weighted step-level scoring.

Opus 91.3% > GPT 88.3% > Gemini 84.1% > DeepSeek 81.2%

Ready to Bring Your Project to Life?

Let our expert team evaluate your needs and provide tailored solution recommendations.

AI Research | Olivenet - Entropy Hunter & ThermoQA | Olivenet