Using Machine Learning to Predict the Viability of
Metal Hydride based Hydrogen Compressors
ABSTRACT
Machine learning
techniques have become
popular methods in the
field of Material Science
for narrowing the search
space for materials with
desired properties for
experimentation and study.
Using the Material-
Agnostic Platform for
Informatics and
Exploration (MAGPIE), a
composition-based dataset
featurizer by Ward et al, a
reproducible machine
learning model was
developed for predicting
the enthalpy of
hydrogenation of metal
hydrides as possible
hydrogen compressors for
hydrogen storage.
Experimentation and
synthesis of viable metal
hydride-based hydrogen
compressors may enable
far more efficient
hydrogen storage and will
allow the development of
hydrogen powered
vehicles and refueling
stations in the near future
GOAL
This project attempts to
reproduce and improve
upon the results of a
previously developed
model which applied
Random Forest Regression
to predict material
enthalpies. While also
replicating such a model,
other machine learning
algorithms are also tested
to seek more promising
results.
NIST Materials Measurement Laboratory, Materials Measurement Science Division
Glen Joy, Dr. Jason Hattrick-Simpers
INTRODUCTION
Hydrogen storage materials are a
method of hydrogen compression
which may be a suitable
alternative to traditional methods.
The Department of Energy has
released techno-economic
efficiency targets for hydrogen
delivery which this study will abide
by in searching for a viable
storage material.
This study focuses on the search
for metal hydride (MH)
compounds for hydrogen storage.
To meet the targets above, a MH
material must have an enthalpy on
the order of 22 kJ molH2-1 ,
reaching pressures of 875 bar at
temperatures on the order of 120
degrees Celsius. To compensate
for model uncertainty, we will
consider enthalpies on the domain
15 ≤ ΔH ≤ 40
DATASET CLEANING
The training dataset used in this
study was the “Hydrogen Storage
Materials Database” from the
Department of Energy's Fuel Cell
Technologies Office. Although the
DOE provided many dataset
filtering options, all dataset
cleaning was done locally.
Non-reversible metal alloys
Stage 1 removed
(>1815)
Compounds with unknown
Stage 2 enthalpies removed
(>545)
Compounds where entropy cannot
Stage 3 be calculated were removed
(>503)
A holdout dataset with 50
compounsd of known enthalpy was
also used to test model prediction
accuracy.
RUNNING MODELS
Hyperparameters are algorithm
values that can be tuned to create a
more accurate model for a particular
dataset. Hyperparameter
optimization was performed using the
Python machine learning library
‘Scikit-Learn’ as well as the WEKA
Experimenter Module. After
performing hyperparameter
optimization, the models were rerun
on the training set with the results
shown below. Model performance was
tested on the basis of 10-Fold Cross
Validation statistics and accuracy to a
set of holdout data.
Actual enthalpies are on the
represented on the x-axis and
predicted enthalpies on the y-axis.
A totally accurate model would have
each data point along the lines on the
parity plots.
ANALYSIS
Based on the results, it appears
that of the three tested models K*
performed the best.
K* yielded the lowest Mean
Absolute Error (MAE) and Mean
Relative Error (MRE) of every
model including Random Forest
.Of the tested models it had the
highest Pearson’s, Spearman’s, and
Kendall’s correlation however, it
did not have higher correlation
coefficients than Random Forest.
Although yielding impressive
statistics, it cannot be said that K*
performed better than Random
Forest but they are certainly
similar if not comparable.
Random Gaussian
KSTAR IBk
Forest Processes
MAE 4.1652 3.1496 4.981 5.5904
MRE 0.01468 0.00326 0.13356 0.04592
R 0.7801 0.7101 0.6444 0.5836
RHO 0.7766 0.6972 0.6194 0.6749
TAU 0.5999 0.5280 0.4532 0.4973
K* was the most computationally
expensive model of every model in
terms of runtime, taking multiple
hours to complete while other
models executed in less than 30
minutes.
CONCLUSIONS
Random Forest Regression was
found to be the most accurate and
efficient model in predicting
hydrogenation enthalpies in this
study while models such as K*
performed astonishingly similar.
Although K* performed well the
results were only comparable to
Random Forest and were not
significant enough to possibly
consider for substitution however,
this does not exclude the use of K*
for future projects. Continued
study of hydrogen storage
materials such as metal hydrides
through various machine learning
models may provide insight into
plausible hydrogen storage
techniques and may bring us closer
to making hydrogen powered
vehicles a reality.