Machine Learning applied to soil compaction.

In this article, we will use Machine Learning techniques to optimize soil compaction.

Soil contains solid, liquid and gaseous components and forms through weathering and sedimentation. Compaction applies a load to reduce air-filled voids and increase soil density.

The compaction curve consists of a curve that is constructed after marking the pairs of dry-density and moisture values of the sample. With a minimum of 5 points, a curve is drawn by hand that can correlate all these points and the maximum dry density and the optimum moisture content at which this soil can be compacted are obtained.

Soil classification and compaction properties are determined using different tests. This project seeks a statistical relationship that can estimate compaction results from soil characteristics.

Problem description

For each kilometer of road, 5 test pits are required, of which two tests must be submitted:

Soil classification
Compaction

In practice, due to the economic resources and logistics involved, only 5 classification tests and 2 compaction tests are carried out and the results are interpolated.

Objective

Determine using Machine Learning techniques a correlation between soil classification and soil compaction, referring to the maximum dry density and optimum moisture content value, to calculate the moisture content at which a certain section of road will be compacted.

Model selection

Because the result is known and the problem is numerical, the models used were:

Linear regression
Multiple linear regression
Polynomial regression

Techniques implemented

Dimension reduction

We removed the columns that did not provide relevant information for this analysis: ‘Pozo’, ‘Lado’, ‘Prf o (m)’, ‘Prf f (m)’, ‘% Hum’, ‘ST >3’, ‘3’, ‘AASTHO’, ‘#’, ‘SUCS’, ‘100%’, ‘97%’, ‘95%’, ‘100%.1’, ‘97%.1’, ‘95%.1’.

Missing values

The LL column had 30 missing values. After examining its distribution with a histogram, we chose to impute them using the median, 32.2.

Graphics

All the necessary graphs were made to understand and interpret the problem.

Linear relationship (with sklearn)

2nd degree polynomial relationship (with numpy)

3rd degree polynomial relationship (with numpy)

Multiple linear relationship (with sklearn)

Multiple linear relationship (with sns) where the 14 parameters are related to each other and their relationship with the optimum moisture content.

Model evaluation

D max vs H opt

Linear regression:

2nd degree polynomial regression:

3rd degree polynomial regression:

Feature vs H opt

Linear regression:

Analysis of results

D max vs H opt

It can be seen that it is the polynomial regression that has a correlation coefficient closest to 1 and in the graph the one that best fits, which is why it will be the regression adopted

Feature vs H opt

A correlation coefficient value close to 1 can be observed and in the sns graphs the vast majority are linear relationships, so the multiple linear model achieved is accepted.

Conclusion

A correlation was determined between soil classification and soil compaction, to calculate the moisture content at which a certain stretch of road will be compacted:

H opt = 0.117*(LL) + 0.0212*(LP) — 0.007*(2) + 0.010*(1 ½) — 0.004*(1) — 0.001*(¾) + 0.017*(⅜) — 0.001*(no 4) — 0.052*(no 10) +
+ 0.054*(no 40) + 0.036*(no 200) — 0.043*(% Grava) + 0.007*(% Arena) + 0.036*(% Finos) + 3.779