The Goldsmith LabFirst-Principles Modeling of Catalysts and Materials
University of Michigan – Ann Arbor, Department of Chemical Engineering
The Goldsmith LabMethods B: Data science and machine
learning for materials
Pollution abatment
Methods A: First-principles modeling
and molecular simulation techniques
Theme 1: Catalysis for Air and Water
Pollution Reduction
Theme 2: Machine Learning to
Accelerate Catalyst Design
Interpretable Machine Learning for Catalysis
Theme 3: Catalysis for Renewable
Energy Generation, Use, and Storage
Density functional theory Minima and saddle
search algorithms
Slab modelsAb initio thermodynamics
and rate theories
[1]. J. A. Esterhuizen, B.R. Goldsmith, S. Linic. Chem (2020).
[2]. H. Wang et al. Nature Comm. 10 (2019).
➢ The objective of this research is to develop
physically transparent and accurate structure-
property models for understanding catalysts.
Rh3S4(100) slab model
Cutting-edge computational techniques yield an accurate description of catalyst
and material electronic and geometric properties under realistic conditions
[1]. J. P. Perdew and K. Schmidt, AIP Conf. Proc., 577 (2001); [2]. M. K. Sabbe et al., Catal. Sci. Tech. 2 (2012); [3]. B. R. Goldsmith, J. Florian et al.,
Phys. Rev. Mater. 3 (2019); [4]. K. Reuter and M. Scheffler, Phys. Rev. B 65 (2002); [5]. K. Reuter et al., Phys. Rev. Lett. 93 (2004); [6]. B. Peters, J.
Phys. Chem. B 119 (2015).
[1-2]
Rh(CO)2 / TiO2(101)Model the system at realistic
chemical potentialsReaction rates
200 K
Ρ = 68%
26%
1%
6%[3]
[5]
[6]
Ab initio molecular dynamics
Combining first-principles (electronic-structure theory) modeling
and data science to understand catalysts and materials
Goldsmith Lab (Fall 2020)
• 5 PhD students
• 2 MS students
• 4 Undergraduate students
Metal and Bimetallic Catalysts for
Bio-Oil Hydrogenation
We recently reported the use of isolated
Pt1 atoms on ceria as “seeds” to develop a
Pt oxide nanocluster for CO oxidation, which
is well-represented by a Pt8O14 model
cluster that retains 100% metal dispersion.
The Pt atom in the nanocluster is 100–1000
times more active than their single-atom
Pt1/CeO2 parent in catalyzing the low-
temperature CO oxidation under oxygen-
rich conditions.[2]
We plan to accelerate these methods
more with ML.[1]. J.-X. Liu, D. Richards, N. Singh, and B. R. Goldsmith. ACS Catal. (2019).
[2]. N. Singh and B. R. Goldsmith. ACS Catal. (2020).
Data analytics applied to catalyst data offers opportunities to advance discovery
We develop and apply machine learning approaches to uncover catalytic insights
Explainable Boosting Machines
➢ Find interpretable global models of a
target property in catalyst data
EBM model predictions easily visualizable through
‘feature shapes’
Active learning/high-throughput
Machine learning applied to alloys yields insights into
the effect of the number of d-electrons in the ligand
metal for various adsorbates.[1]
Pt Surface
Ligand
Exploring alloy-induced ligand and strain effects
on adsorption properties of alloys. In
collaboration with Prof. Suljo Linic’s lab.[1]
Strain effect Ligand effect
Single Atom and Nanocluster Catalysis for CO2 reduction
➢ CO2 can be reduced via H2 to produce either
methane (CH4) or carbon monoxide (CO),
depending on the presence of nanoclusters
or single atoms.
➢ What impact on catalytic activity and
selectivity can be seen from varying the
metal nanocluster size and support surface?
➢ Aqueous nitrate (NO3–), a major water
pollutant, can be remediated with
electrocatalytic nitrate reduction.
➢ Metal alloys can perform this reaction with
higher activity (higher turnover frequency,
TOF) than their pure-metal counterparts.
➢ What alloy compositions make the most
active and selective catalyst?
➢ In collaboration with Prof. Nirala Singh’s
lab (UofM ChE).
Redox Flow Batteries for
Large-Scale Energy Storage
➢ Redox flow batteries are used to match
power grid supply to demand, which is
increasingly relevant as we transition to
intermittent renewable energy sources.
PtxRuy alloys for electrocatalytic nitrate reduction
● = Pt, ● = Ru
➢ Use machine learning (ML) to rapidly and
accurately predict catalyst figures of merit.
➢ Strategically train ML model reduce
number of costly DFT calculations.
[3]. Akinola, J.; Barth, I.; Goldsmith, B. R.; Singh, N. ACS Catal. (2020).
➢ Biomass-derived molecules can be
upgraded to fuels and industrially relevant
chemicals using aqueous-phase
electrocatalytic hydrogenation driven with
renewable electricity
Phenol Cyclohexanol
[2]
[2] Settles, B. Active Learning. (Morgan & Claypool Publishers, 2012), p. 8.
Active learning strategy enables rapid screening of
catalyst descriptors, such as adsorption energies
[3]
[4]
[3]
[1]
➢ We recently reported an anion bridging
mechanism for the V2+/V3+ redox
reaction on glassy carbon electrodes
for vanadium redox flow batteries.[1]
➢ We elucidated the structures and free
energies of Ce3+ and Ce4+ in various
electrolytes.[2]
[1]. Agarwal, H.; Florian, J.; Goldsmith, B. R.; Singh, N. ACS Energy Lett. (2019).
Ch
an
ge
in
ad
so
rptio
n e
ne
rgy
Change in ligand d-electrons relative to host metal
[2]. C. Buchannan et al. Inorg Chem. (2020).
[1] B. R. Goldsmith et al. AIChE J. (2018).
[1]
Machine Learning Enabled High-Throughput Evaluation of Catalysts