Here, we propose a suite of ‘top-down’ design and ‘bottom-up’ discovery
of porous metal oxides metal oxides and anticipate the formation of
extended materials based upon these molecular building blocks, see
Figure 4. Leveraging the decades-long expertise of the broad
computational material science community and the large amount of data
available at Density Functional Theory (DFT) level will be beneficial to
obtain reliable description of POMs electronic structure, which can be
used to generate the database obtained by exploring the chemical space
of POMzite materials.[25] This initial database
will be explored and refined using principal component analysis that
will be used to establish a new bottom-up method. Top-down methods will
rely on analysis of the structural database, and the development of deep
neural network produce new synthetic routes by developing theory from
aggregated observations using Reinforced
Learning
(RL) algorithms.
Figure 4. The estimated accessible chemical space generated by expanding
the POMzite family.
The general objective is to develop a new conceptual approach for the
fabrication of functional nano-molecules and adaptive materials.
Proposing to explore the hypothesis of first predicting and then
synthesizing new POMzite materials. The ideal outcome of the research
proposed here is to predict new POM all-inorganic framework materials,
followed by experimental realization by our collaborators. By
theoretically expanding the acessible chemical space generated by the
derivatization of polyoxometalate (POM) clusters, which enables their
assembly into a range of frameworks by use of inorganic
linkers.[2] These robust all-inorganic frameworks
are made using metal-ion linkers (Co, Mn, Ni, Ag), which combine
molecular synthetic control without the need for organic
components.[1] Later generating the heterogeneous
database of POMzite materials, 1.6 107 estimated
structures, see Figure 4. For instance, so far we have just explored 4
transition metals as linkers, out of 30 possibilites
We have the tools to understand POMzites at a molecular level and
understand the interactions between the electrons of atoms bound in
molecules using ‘first principle’ techniques DFT algorithms. By mapping
their electronic structure by modelling, using an ab init o
packages with basis sets and functionals for transition metal clusters.
The cyclic heteropolyanion
[P8W48O184]40−(abbreviated as {P8W48}) is the
current building block for the construction of porous framework
materials. Preliminary gas-phase calculations confirmed the stability of
{P8W48} as an initial structural
motif, [26] derivatives such as the Se equivalent
are also stable. Theoretically derivate other cyclic POMs with different
heteroatoms (As, Sb, S, Te) and asses their potential to aggregate into
porous structures. We are also interested in understanding the binding
sites of the POMzite materials, this will only be possible by describing
their molecular structure. This basic computational setup will provide
with data (e.g. molecular charge, optimized geometry, HLgaps, etc.) to
start the initial layer of a neural network.
Given the size and the characteristics of their unique porous
environment, we will need macroscale modelling of the POMzite
architectures. To advance this area, we need to combine experimental
input with computational modelling to imagine a wide range of different
rearrangements, and to create a blueprint for building these
materials.[27] Aiming to understand the
connectivity and the stability of
{P8W48} rings in POMzite structures
and predict new motifs from the results, as well as being able to
anticipate new favourable structures with different linker modes. Given
the fact that some POMzites have just been observed with Co or Ni, e.g.
POMzite-5, by modelling we will be able to substitute those linkers with
another TM. We use simple topological methods such as stochastic
modelling to make predictions of the different binding
sites.[28] We will be able to predict new bulk
physical properties from assemblies of these building blocks e.g.
electropotential, acidity, optical properties, electronic storage. Once
we have fully understood the chemical system and created a database of
all the molecular architectures, we would like to predict their physical
properties.
The data generated in this project will provide a greater insight in the
binding site and the resulting in the crude heteregoneous framework
materials database, with an estimate of 1.6 107structures. The curation of this database will require initial data
pre-processing of the POMzite geometries, checking its format and
presenting them in an understandable format. This will be the ideal
training set to use an end-to-end Reniforcement Learning (RL) approach
to investigate whether is possible to predict new and stable POMzite
materials. We propose to use a modular behaviour-based reinforcement
architecture that will start training neural networks with a known
solution,[29] leveraging on knowledge of known,
and/or predicted to be stable, POMzite materials. Classically a learning
algorithm training will split the dataset into 3 sets: training,
validation, and testing datasets. The first one is used in the learning
process, where the model parameters are obtained. Once we obtain an
optimal set of parameters, the test set is used to assess the
performance of the model. If the obtained model is unsuccessful, the
previous steps are repeated with improved data selection,
representation, transformation, sampling, and removing outliers, or by
changing the algorithm altogether.[10]