We collected measurements with two thermopile sensors with relative values only (no heat…
- User
We collected measurements with two thermopile sensors with relative values only (no heat power measurement). The gas mixture was H2, N2 and CO2. We can see a linear relationship between thermal diffusivity and our sensor data. We see similar correlation with thermal conductivity. However, we believe this is due to the correlation already present in the data. We see a straigth line if we plot both gas properties together. If we wanted to know if our sensors can observe or estimate thermal conductivity, we would need to use gas mixtures, where there is a more complicated relationship between the gas properties, correct?
Assistant
Yes, you're absolutely right! Your current gas mixture (H₂, N₂, CO₂) creates an essentially 1D manifold in property space because you're varying only the mixture proportions. Since thermal conductivity and thermal diffusivity are highly correlated along this 1D path, you cannot distinguish their individual effects on your sensor readings.
The Core Problem
Your sensors respond to some combination of gas properties, but with your current data:
- You can estimate a linear combination of properties (like "effective thermal property")
- You cannot uniquely identify thermal conductivity vs. thermal diffusivity separately
- The correlation means there are infinitely many combinations that fit your data equally well
What You Need: Break the Correlation
To test if your sensors can actually observe thermal conductivity specifically, you need gas mixtures that create a 2D or higher-dimensional manifold in property space. This means:
Option 1: Add More Gas Components
Instead of just H₂/N₂/CO₂, include gases with very different properties:
- Helium (He): Extremely high thermal conductivity, low density
- Argon (Ar): Moderate conductivity, higher molecular weight
- Methane (CH₄): Different molecular structure than your current gases
This creates more diverse property combinations that can break the linear relationship.
Option 2: Vary Temperature and Pressure
If possible, measure at different:
- Temperatures: Properties change differently with T (k ∝ T^α, α varies by gas)
- Pressures: Density changes affect properties differently
Option 3: Use Binary Mixtures Systematically
Create mixtures that span a wider property space:
Pure gases: H₂, He, N₂, Ar, CO₂ Binary mixes: H₂+CO₂, He+N₂, H₂+Ar, etc.Concrete Experimental Design
1. Property Space Visualization
First, plot your current data:
- X-axis: Thermal conductivity (λ)
- Y-axis: Thermal diffusivity (α)
- Each point represents one mixture composition
If all points fall on/near a straight line, you need more diversity.
2. Design Target: Cover 2D Property Space
Aim for gas mixtures that create a scatter in the λ-α plane, not a line. Example combinations:
- High λ, low α: He + CO₂ mixtures
- Low λ, high α: H₂ + Ar mixtures
- Medium λ, medium α: Your current mixes
3. Test Protocol
Once you have diverse gas mixtures:
Step 1: Train a model to predict both λ and α from sensor data
[Sensor₁, Sensor₂] → [λ, α]Step 2: Cross-validation with holdout regions
- Hold out entire regions of property space
- Test if your model generalizes to unseen λ-α combinations
Step 3: Ambiguity analysis
- For each sensor reading, find all similar readings in your dataset
- Check if they correspond to similar or different (λ,α) pairs
- High ambiguity = poor observability
Step 4: Sensitivity analysis
- Compute Jacobian: J = ∂[Sensor₁,Sensor₂]/∂[λ,α]
- Check condition number of J
- Poor conditioning = difficulty distinguishing λ from α
Quick Check with Current Data
Before running new experiments, analyze your existing data:
- Correlation coefficient between λ and α in your current dataset
- Condition number of the [λ, α] data matrix
- Principal component analysis: what fraction of variance is in the first PC?
If >95% of variance is in the first PC, you definitely need more diverse mixtures.
Alternative Approach: Sensor Physics
Consider whether your sensors are fundamentally sensitive to:
- Thermal conductivity (heat transfer rate)
- Thermal diffusivity (speed of thermal wave propagation)
- Specific heat (thermal mass)
- Some combination
The underlying physics might make one property more "observable" than others, regardless of your gas mixtures.
Would you be able to share a plot of your current λ vs. α data? That would help determine exactly how correlated your properties are and what kind of gas mixtures would best break that correlation.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.