Accelerating f-element separation with machine learning

A grand challenge in chemical separation currently stands in the way of the development of abundant, affordable, and reliable forms of clean energy.

Our central scientific mission is to accelerate f­ element separation science design using an integrated data-driven autonomous discovery loop that explores the vast space of separation chemistries with minimal human bias, in a way that provides fundamental molecular-level understanding of the separation process. Toward this end, this proposal concentrates on separation of critical rare earth materials from one another and from actinides relevant to nuclear energy generation, with a particular focus on +3 f-elements, specifically Neodymium (Nd), Europium (Eu), Terbium (Tb), Dysprosium (Dy), Holmium (Ho), Yttrium (Y), Americium (Am) and Curium (Cm). We will employ a multidisciplinary approach that combines high-throughput computations and experiments integrated together with modern data-science and artificial intelligence approaches to shed light on the fundamental understandings to enable a low-carbon future. 

Clean Energy Future

A grand challenge in chemical separation currently stands in the way of the development of abundant, affordable, and reliable forms of clean energy. Critical 4f rare earths and minor Sf actinides are all +3 f-elements with similar chemistries, but often very different applications. Separating these elements from one another in high purity and in high yield is essential to the success of the DOE Clean Energy Initiative, but separation is extremely difficult given their nearly indistinguishable chemical characteristics.

Accelerated Separation Discovery

ML for Separation: Integrating machine learning, simulations, and robotics for chemical separation and discovery.

Our central scientific mission is to accelerate f­ element separation science design using an integrated data-driven autonomous discovery loop that explores the vast space of separation chemistries with minimal human bias, in a way that provides fundamental molecular-level understanding of the separation process. Toward this end, this proposal concentrates on separation of critical rare earth materials from one another and from actinides relevant to nuclear energy generation, with a particular focus on +3 f-elements, specifically Nd, Eu, Tb, Dy, Ho, Y, Am, and Cm. We will employ a multidisciplinary approach that combines high-throughput computations and experiments integrated together with modern data-science approaches to shed light on the fundamental understandings to enable a low-carbon future.

Robotic lab equipment controlled by SeparationML machine learning software.

Automating Separations: LANL Super Separator

  • Makes it easier to develop a new process.
  • Makes it easier to optimize an operational process with the confines of an existing safety envelope.
  • Automation increases throughput and minimizes human error.
  • Commissioned for radioactive actinides.
Chemical separation techniques: solvent extraction, extraction chromatography, and selective precipitation.
Five steps for SeparationML chemical extraction with database integration and high-throughput experiments.

Design Areas:

Thrust 1: Design and screen highly selective extractants by controlling coordination enviornment
High-throughput workflow and computation Architechtor, density functional theory, density functional tight binding theory, pyrion, machine learning model

Thrust 2: Navigate the vast f-element separation landscape with data-driven high throughput robotics
High-throughput robotic experiment, the Super Separator, Bayesian optimization 

Thrust 3: Understand temporal changes in structure and speciation at the molecular level
System identification, Chemical reaction network, molecular dynamics simulations, experimental spectroscopic validation

Software Packages

Periodic table highlighting alkali/alkaline earth metals, post-transition metals, metalloids, nonmetals, lanthanides and actinides.

Architector

Architector is a state-of-the art software for performing in-silico 3D design of coordination complexes including any chemically accessible metal-ligand combinations. Architector leverages metal-center symmetry, interatomic force fields, and tight binding methods to build many possible 3D conformers from minimal 2D inputs including metal oxidation and spin state. It was demonstrated on a set of over 6,000 coordination complexes spanning the periodic table with demonstrated quantitative agreement with experimentally observed structures.

Overall, Architector represents a transformative step towards cross-periodic table computational design of metal complex chemistry. It was originally published in Nature communications, has been highlighted in the ACS publication CE&N news, and was highlighted in Nature Computational Sciences. It is readily available as open-source software via GitHub or conda where it been installed over 5,000 times as of 4/2024.

The process used by Minervachem, induce subgraphs, regress, combine and adjust.

Minervachem

Minervachem is a tool for fast, accurate, and interpretable molecular property prediction with uncertainty quantification based on molecular graphlet fingerprints. It provides both rdkit and parallelizable sklearn-style dataset transformer interfaces for creating molecular fingerprints, and a hierarchical linear regressor for accurate prediction on large chemical datasets.

It also provides methods for projecting learned model coefficients onto atoms or bonds in molecular graphs for intuitive prediction visualization.  It also provides an implementation of pairwise difference regression (PADRE), which both improves accuracy and provides uncertainty for an arbitrary base regressor.

High-throughput simulation for chemistry via pyiron

The pyiron workflow framework offers a robust platform for developing complex simulation workflows. Originally crafted for computational materials science, the SeparationML project team has adeptly extended pyiron to encompass chemistry. This expansion includes interfaces to various simulation codes, ranging from ab-initio codes tailored for actinide chemistry to tight-binding models and interatomic simulation codes. The capability to integrate these diverse levels of theory with in-house machine learning initiatives is crucial for automating the separation project and facilitating high-throughput screening in separation chemistry.

Moreover, the pyiron framework has been enhanced to support scaling up to leadership computing resources. Within pyiron, simulation workflows are constructed by combining specialized Python objects—akin to building blocks—with pyiron managing communication with high-performance computing (HPC) resources and data storage. This synergy enables the rapid prototyping of new simulation methods and their efficient scaling for separation chemistry with minimal overhead. 

Media Coverage

Chemists have a new tool to predict 3D structures of f-block organometallics image

Chemists have a new tool to predict 3D structures of f-block organometallics

An application called Architector could help scientists separate valuable metals from nuclear waste

Read More
Building molecules across the periodic table image

Building molecules across the periodic table

Computational tools that successfully generate three-dimensional (3D) molecular configurations have become increasingly available for a variety of applications...

Read More

PERSONNEL

SENIOR PIs

Ping Yang

Ping Yang is the Deputy Director of the G.T. Seaborg Institute for Transactinium Science and a Staff Scientist in the Physics and Chemistry of Materials Group of the Theoretical Division at Los Alamos National Laboratory. She has extensive experience in computational approaches to modeling electronic structure and reactivity of actinides, separation, surface chemistry, and nanomaterials in solution environments. Her research focuses on understanding fundamental electronic structures, spectroscopic properties, reactivity, and dynamical behaviors of f-element chemistry and materials that are crucial for energy security.  She has broad interests in both applications of current high-performance computing frameworks and development of new computational methods for long time-scale simulations of complex f-element systems. 

Stosh Kozimor

Dr. Kozimor is a staff scientist in the Inorganic, Isotope, and Actinide Chemistry group (C-IIAC), where he is heavily involved in the isotope production team.  He focus on the study of the unique properties and problems associated with radioactive elements, including synthetic expertise and spectroscopic expertise for f-elements. He leads a laboratory-based team in conducting X-ray spectroscopy measurements on radioactive samples at synchrotron facilities since 2006.In addition, Stosh serves as the Los Alamos National Laboratory BES HEC project PI and maintains a fully functional transuranic laboratory (the "Alpha Wing"). Stosh has a wide background in inorganic synthesis, spectroscopy, and radiochemistry with his work at the Lab branching out into studies of actinides (in particular transuranics). Of particular relevance, Stosh spent four years working on the C-NR radiochemistry, where his technical focus was center on nuclear forensics. He was additional the P.I. for the DTRA supported Nuclear Forensic Undergraduate Summer School.

Joshua Schrier

Joshua Schrier is a physical chemist interested in using computers to accelerate the discovery of new materials, by using a combination of physics-based simulations, cheminformatics, machine learning, and automated experimentation. He is the Kim B. and Stephen E. Bepler Professor of Chemistry at Fordham University in New York City.  Prior to joining Fordham in 2018, he was on the faculty at Haverford College, and a Luis W. Alvarez computational sciences postdoctoral fellow at Lawrence Berkeley National Laboratory. As a faculty member, he has received awards include the Dreyfus Teacher-Scholar, U.S. Department of Energy Visiting Faculty, and Fulbright scholar awards.

Sara Adelman

Adelman's diverse educational background, including a PhD in Chemistry from Michigan State University under Professor Jim McCusker, and her professional experiences have led her to a career at Los Alamos National Laboratory. There, she focuses on complex scientific questions vital for national security. During her time at MSU, she worked on developing earth-abundant dyes for solar cells and analyzed their electronic and physical properties. As a post-doc at the Laboratory, under Drs. Stosh Kozimor, Veronika Mocko, and Benjamin Stein, she developed a robotic system for automated chemical processes in a radiological lab. Now, as a scientist, she enhances automated systems for pharmaceutical isotope production and nuclear energy applications, collaborating with data scientists to improve separations and train machine learning algorithms. 

Enrique R. Batista

Enrique R. Batista, received a B.S. in physics, with honors from the University of Buenos Aires, Argentina, M.S. and Ph.D. in physics from the U. of Washington. Post Ph.D. he worked as a postdoctoral fellow at Columbia University, in the Environmental Molecular Sciences Institute, funded by the NSF.  He joined Los Alamos National Laboratory in 2003 as a post-doctoral fellow at the Theoretical Chemistry and Molecular Physics group and a staff scientist in 2005.  Since 2016 Batista is Deputy Center Director of the Center for Nonlinear Studies (CNLS).  His research areas of interest in the computational study of actinide chemistry in gas phase and in solution, spectroscopy, coordination, and reactivity, and on electronic structure studies of transition metal complexes with application to homogeneous catalytic processes.  His studies also extend to solid-state systems and surface chemistry problems. Batista works with, and towards the development of, efficient computational methodologies for atomistic simulations and electronic structure approaches. 

Nicholas Lubbers

Nicholas Lubbers is a computation scientist with Los Alamos National Laboratory’s Information Sciences group. His background is in statistical physics, with a Ph.D. from Boston University under Professor William Klein. He had a postdoctoral appointment with the Center for Nonlinear Studies and Theoretical Division at the Lab, and in 2018 the Information Sciences group in the Computer, Computational, and Statistical Sciences division as a staff scientist. He is an organizer and mentor for the Laboratory’s Applied Machine Learning Summer Program, which he lead in the years of 2021-2023.

His research focus is on Machine Learning for Science, with diverse applications to Materials Science, Seismology, Fluid Mechanics, Microbiome analysis, Nanoscale phenomena, and Computational Chemistry. He has particular interest in physics-informed algorithm design, uncertainty quantification and active learning, and meta-algorithms. His mission is to contribute to science and technology through innovation, mentoring, and open-source software development. 

Danny Perez

Danny is a Scientist IV at Los Alamos National Laboratory. He obtained a Ph.D. in Physics from Université de Montréal in 2006. He then joined the Laboratory as a Director’s postdoctoral fellow in 2007. His research centers around the development of long-time atomistic simulation methods that are designed to explore systems that evolve through sequences of rare events. He then implements these methods in high performance simulation codes, and applied them to a range of problems of relevance to energy applications, such as nuclear fusion or fission. Over the last few years, Danny has been especially interested in developing strategies for the optimal use of exascale computing. He is currently the PI of the EXAALT project funded by DOE’s Exascale Computing Project where his team develops novel ultra-scalable approaches for materials simulations, including long-timescale molecular dynamics, optimization for GPUs, and high-throughput large-scale workflows. 

Early Career

Zafer Acar

Zafer Acar is a Ph.D. candidate at Michigan Technological University, holding a Bachelor of Science degree in Astrophysics and a Master's degree in Physics. Zafer has served as a software engineer and database administrator for financial institutions in Los Angeles. He is an enthusiastic advocate of big data, driven by a profound interest in Data Science, Machine Learning, and Artificial Intelligence. His insatiable curiosity and dedication have led him to conduct research in prestigious institutions such as the European Nuclear Research Center, Princeton University, UC Berkeley, the United States Navy, and the Los Alamos National Laboratory. His research interests encompass Machine Learning, Natural Language Processing, Computer Vision, Computational and Nuclear Chemistry, and Solvent Separation.

Alexander Brown

I was first introduced and motivated to pursue science by a collection of 20th century science fiction writers. While I eventually concluded that faster-than light travel wasn’t possible yet and mathematical psychoanalysis couldn’t predict the future, I found a place at the interface of physical inorganic chemistry, materials science, and biology where challenges related to human health and sustainability can be addressed from fundamental physical principles. While these areas appear disparate, they are driven by a common thread – by uncovering molecular and electronic structure properties through synthesis, physical characterization, and theory, rationally design materials can be constructed. After earning a B.Sc. in biochemistry from Bridgewater State University I began my doctoral studies at Brown University working under Prof. Jerome Robinson. Here, my interest in developing structure-function relationships led to several projects in the fields of radiopharmaceutical development, oxygen reduction catalysis and ring opening polymerization reactions where an atomistic understanding of the chemical system is critical for achieving desirable results. These studies have branched into my pursuits at LANL where I am working as a DOE SCGSR fellow utilizing a combination of high-throughput simulation and synthetic modeling to evaluate the impact of chelate binding on the electronic structure and redox potentials of actinide elements for application to separations technologies. 

Jiyoung Lee

I was born and grew up in Ulsan, South Korea before moving to Seoul for my undergraduate. I studied chemistry at Ewha Womans University where I found my strong interest in physical chemistry. After I finished my Bachelor’s, I came to US and did Master’s in computational chemistry at Iowa State University. I was part of a project to improve interoperability of quantum chemistry packages and it led me to pursue interdisciplinary research where I am currently being trained at University of Texas at Austin. My research interest at UT was to apply machine learning technique to accelerate searching local minima and saddle points.  

Thomas Summers

A native of Tennessee, Thomas Summers received his B.S. in biochemistry from Christian Brothers University in Memphis. He continued on to receive his M.S. and Ph.D. in chemistry from the University of Memphis, his research focusing on modeling enzyme-catalyzed reactions and protein inter-residue interaction networks. He then joined the University of Nevada, Reno as a postdoctoral researcher for two years, where he worked on characterizing the solution structures of lanthanide-ligand complexes through molecular simulation and extended X-ray absorption fine structure spectroscopy. He recently joined the Theoretical Division at Los Alamos National Laboratory as a postdoctoral researcher where he continues to explore the coordination structures formed by f-elements using high-throughput computational modeling. 

Michael Taylor

Michael Taylor is a chemical engineer interested in physics-based simulations, cheminformatics, and machine learning to discover new materials and chemistries.   He is a staff scientist in the Theoretical Division at Los Alamos National Laboratory.  Prior to a recently completed postdoc at the Laboratory he was a postdoc at MIT. Prior to MIT he completed his PhD in chemical engineering at the University of Pittsburgh as an NSF graduate research fellow (GRFP).   

Mike Tynes

Mike joined Separation ML as a Post-Masers fellow in Los Alamos National Laboratory’s T-1, after completing a masters degree in data science from Fordham University. As part of Separation ML Mike has developed algorithms and tools for machine learning in molecular property prediction which improve model accuracy, interpretability, and provide uncertainty quantification. These methods are available in our open source package minervachem. Mike also contributed to the Lab’s cinema scientific visualization tool improving support for categorical variables and for organic/inorganic chemical data. Mike is now a PhD student at the University of Chicago Department of Computer Science, advised by Dr. Ian Foster.

Yufei Wang

I am a postdoc working on the application of an automated system in high-throughput separations. Before I joined Los Alamos National Laboratory, I was a Ph.D. student at UC Berkeley where I performed research on the separations of lanthanides and actinides by liquid-liquid extraction. My interests include badminton, swimming, and skiing. 

Alumni

Brian Arko
Dr. Daniel Burrill
Dr. Jan Janssen
Rebecca Li 

PUBLICATIONS