The BOS-TMC Dataset: DFT Properties of 159k Experimentally Characterized Transition Metal Complexes Spanning Multiple Charge and Spin States

This paper introduces the BOS-TMC dataset, a comprehensive collection of over 2.9 million DFT properties for 159,000 experimentally characterized transition metal complexes across multiple charge and spin states, designed to serve as a high-fidelity foundation for machine learning, DFT benchmarking, and chemical exploration.

The Gist

Imagine you are trying to teach a robot how to understand the chemistry of the future. To do this, the robot needs a massive library of examples. It needs to see millions of different molecules, understand how they behave, and learn the rules that govern their energy and stability.

For a long time, this library was missing a crucial section: Transition Metal Complexes. These are molecules with a metal atom (like iron, copper, or gold) at the center, surrounded by other atoms. They are the "workhorses" of chemistry, used in everything from making medicines to creating solar panels. But they are notoriously difficult to study because they are messy, can exist in many different "moods" (spin states), and often carry electrical charges.

This paper introduces BOS-TMC, a massive new digital library designed to fix that problem. Here is the story of how they built it, explained simply.

1. The Problem: The "Messy Attic"

Think of the Cambridge Structural Database (CSD) as a giant, dusty attic filled with millions of blueprints for these metal molecules. These blueprints were drawn by real scientists using X-ray machines, so they are real, physical structures.

However, the attic is messy:

• Missing Labels: Many blueprints don't say what the total electrical charge of the molecule is.
• Multiple Personalities: A single molecule can change its "personality" (spin state). It can be calm (low-spin), energetic (intermediate-spin), or wild (high-spin). Old datasets mostly ignored these different personalities.
• Broken Blueprints: Some blueprints were just sketches that didn't match the real chemistry.

2. The Solution: The "Digital Renovation Crew"

The authors (a team from MIT and the Technical University of Munich) acted like a high-tech renovation crew. They went into the attic and did three main things:

A. Cleaning and Labeling (Data Curation)
They wrote a smart computer program to go through 299,000 blueprints. They threw out the broken ones, fixed the missing hydrogen atoms (the tiny "glue" atoms), and used a clever math trick to figure out the electrical charge of every single molecule.

• The Result: They ended up with 159,000 clean, verified, real-world metal complexes.

B. Exploring the "Moods" (Spin States)
Most old datasets only looked at the "calm" version of a molecule. But in reality, these molecules can get excited. The team calculated the properties of these molecules in up to three different moods (low, intermediate, and high spin).

• The Result: Instead of 159,000 molecules, they now have 343,800 unique "molecule + mood" combinations. It's like having a photo of a person smiling, frowning, and laughing, rather than just one photo.

C. The "Do Not Touch" Rule (Preserving Reality)
Here is the most important part. Usually, when scientists use computers to study molecules, they let the computer "relax" the structure, moving atoms around to find the perfect theoretical shape.

• The Analogy: Imagine taking a photo of a crumpled piece of paper and then using a computer to smooth it out perfectly. You lose the crumple, which was part of the reality.
• The BOS-TMC Approach: They said, "No!" They kept the heavy atoms exactly where the X-ray machine found them. They only moved the tiny, invisible hydrogen atoms to make the math work. This ensures the data reflects real, physical chemistry, not just a perfect computer fantasy.

3. The Treasure Chest: What's Inside?

Once they had the structures, they ran them through a super-accurate calculator (DFT) to generate a massive list of properties. They didn't just stop at one number; they generated 2.9 million data points.

Think of this as a "Molecular ID Card" for every single entry, containing:

• Energy Levels: How hard is it to steal an electron? (HOMO/LUMO)
• The Gap: How much energy does it take to jump from one state to another?
• The Charge: Where is the electricity concentrated?
• The Magnetism: How does the molecule react to a magnetic field?
• Atomization Energy: How much energy would it take to blow the whole molecule apart into individual atoms?

4. The Stress Test: "Which Calculator is Best?"

The team knew that different computer calculators (called "functionals") give different answers. To test this, they took a smaller sample of 10,000 molecules and ran them through 12 different calculators.

• The Finding: They found that for some molecules, the calculators agreed perfectly. But for others—especially Copper (Cu) and Nickel (Ni) complexes—the calculators disagreed wildly (sometimes by huge amounts).
• Why it matters: This highlights exactly where our current scientific tools are weak. It tells future researchers, "Hey, if you are studying Copper, be careful which calculator you use!"

5. Why Should You Care?

This dataset is a game-changer for two reasons:

1. For AI and Machine Learning: If you want to train an AI to discover new drugs or better batteries, you need high-quality, diverse data. BOS-TMC provides the "diverse diet" of metal chemistry that AI models have been starving for. It covers charged molecules and different spin states that previous datasets ignored.
2. For Scientists: It serves as a "gold standard" benchmark. If a new computer method claims to be better at predicting chemistry, scientists can test it against BOS-TMC to see if it actually works on real, messy, charged, multi-mood molecules.

The Bottom Line

The BOS-TMC dataset is like building a massive, high-definition map of a previously uncharted territory. It doesn't just show the mountains; it shows the valleys, the different weather patterns (spin states), and the electrical storms (charges). By keeping the structures true to their real-world X-ray photos, it gives scientists and AI a reliable foundation to build the next generation of chemical discoveries.

Technical Summary

1. Problem Statement

Transition metal complexes (TMCs) are central to catalysis, materials science, and medicine, but their electronic structures are notoriously difficult to model due to strong electron correlation, multiple accessible spin states, and diverse coordination environments.

• Data Scarcity and Bias: Existing machine learning (ML) datasets for TMCs (e.g., tmQM, tmQMg, OMol) suffer from significant limitations:
  • They often focus on closed-shell systems or specific subsets of metals.
  • They frequently exclude highly charged species ($|q| > 1$), which are common in experimental chemistry but rare in curated datasets.
  • They typically rely on gas-phase DFT geometry optimizations, which can deviate significantly from experimental crystal structures, potentially altering bonding trends and electronic properties.
  • They rarely provide data across multiple spin states (low, intermediate, high) for the same complex, hindering the study of spin-crossover phenomena and ground-state reassignment.
• Methodological Uncertainty: Properties like spin-splitting energies are highly sensitive to the choice of the exchange-correlation (xc) functional in Density Functional Theory (DFT), yet large-scale benchmarks across diverse functionals are lacking.

2. Methodology

The authors curated the BOS-TMC (Boston Open-Shell Transition Metal Complex) dataset through a rigorous, multi-stage workflow:

• Data Curation:

  • Source: Extracted from the Cambridge Structural Database (CSD) (March 2024 release).
  • Filters: Selected mononuclear TMCs with a single transition metal, at least one heavy atom ligand, and no polymeric species.
  • Charge Assignment: Used an iterative unit-cell neutrality algorithm to assign formal charges and oxidation states, ensuring chemical completeness.
  • Hydrogen Addition: Added hydrogens using the CSD API with quality checks to ensure valence satisfaction.
  • Final Set: 159,014 converged complexes (126k unique molecular graphs) spanning charges from -8 to +8.

• Geometry and Electronic Structure Calculations:

  • Geometry Strategy: Unlike prior works, heavy-atom coordinates were fixed to experimental crystallographic positions to preserve realistic bonding environments. Only hydrogen positions were optimized.
  • Spin States: Calculated properties for up to three spin states (Low-Spin, Intermediate-Spin, High-Spin) based on the metal's $d$-electron configuration and oxidation state.
    • 3d metals: Up to 3 states (e.g., singlet, triplet, quintet).
    • 4d/5d metals: Up to 2 states (low/intermediate).
  • Computational Protocol:
    • Optimization: PBE0/def2-SV(P) with D3 dispersion correction in implicit solvent (dielectric $\epsilon=10$) using TeraChem.
    • Single-Point Energies: High-fidelity PBE0/def2-TZVP calculations using Psi4.
  • Properties Reported: Electronic energy, HOMO/LUMO levels, HOMO-LUMO gap, dipole moments, Mulliken/Löwdin partial charges, and spin-state-dependent atomization energies.

• Functional Sensitivity Analysis (manyDFA Set):

  • A subset of >10k TMCs (smallest unique structures preserving metal distribution) was evaluated using 12 different xc functionals spanning "Jacob's ladder" (from semi-local PBE to double hybrids like PWPB95).
  • This allowed for the assessment of functional dependence on properties like spin-splitting energies and atomization energies.

3. Key Contributions

• Scale and Diversity: The dataset contains 159k experimentally characterized TMCs with 343.8k spin-state combinations, totaling over 2.9 million computed properties. It is the largest dataset of open-shell TMCs derived from experimental structures.
• Inclusion of Highly Charged Species: Uniquely includes a significant number of complexes with absolute charges $|q| > 1$ (26k+), which were largely excluded from previous datasets.
• Experimental Fidelity: By preserving experimental heavy-atom coordinates, the dataset avoids artifacts introduced by gas-phase geometry relaxation, offering a more realistic representation of crystal environments.
• Spin-State Coverage: Provides a comprehensive view of low-, intermediate-, and high-spin states, enabling the study of ground-state reassignment and spin-crossover trends.
• Methodological Benchmarking: The "manyDFA" subset provides a critical resource for benchmarking DFT functionals against experimental structures, identifying "hotspots" of high uncertainty.

4. Key Results and Findings

• Property Distributions:
  • HOMO-LUMO Gaps: Highly charged species ($|q| > 1$) exhibit larger average gaps (4.24 eV) compared to neutral species (3.78 eV), though the most strongly charged species show shrinking gaps.
  • Dipole Moments: Highly charged complexes have significantly smaller dipole moments (mean 2.9 D) compared to neutral ones (7.0 D).
  • Atomization Energies: TMCs are generally less stable per atom than organic molecules in the QM9 dataset. Late transition metals (Zn, Cd, Pd, Ag) show the highest sensitivity of atomization energy to spin state.
• Spin-State Dependence:
  • Ground State Reassignment: 20% of the TMCs in the dataset have Intermediate-Spin (IS) or High-Spin (HS) ground states, contrary to the Low-Spin (LS) assumption often made in prior datasets.
  • Property Shifts: Reassigning the ground state from LS to IS/HS causes significant shifts in properties:
    • HOMO-LUMO gaps shift by an average of 1.28 eV (IS) and 0.65 eV (HS).
    • Dipole moments shift by ~0.40 Debye on average.
    • Metal partial charges shift by ~0.10–0.18 a.u.
• Functional Sensitivity:
  • Spin-Splitting Energies (VSSE): Show extreme sensitivity to functional choice, with standard deviations up to 40 kcal/mol across the 12 functionals.
  • Outliers: Cu(II) and Ni(II) square-planar complexes, particularly with mixed-field ligands, exhibit the highest variability across functionals.
  • Grouping: Functionals can be grouped by Hartree-Fock (HF) exchange fraction. Range-separated hybrids and double hybrids show high agreement, while semi-local functionals (PBE) and high-HF functionals (M06-2X) often disagree most strongly.

5. Significance

• Machine Learning Foundation: BOS-TMC provides a high-fidelity, diverse training set for developing ML models capable of predicting properties across a wide range of charges, spins, and coordination geometries, addressing the "open-shell" gap in current ML potentials.
• DFT Benchmarking: The dataset serves as a rigorous benchmark for evaluating and selecting DFT functionals for transition metal chemistry, highlighting specific chemical motifs (e.g., Cu(II) square planar) where standard functionals fail.
• Chemical Discovery: By capturing experimentally realized but computationally under-explored regions (highly charged, open-shell), the dataset facilitates the discovery of new catalysts and materials with specific electronic properties.
• Open Science: All data, including structures, properties, and scripts for charge assignment and atomization energy calculation, are publicly available via Zenodo, promoting reproducibility and further exploration.

In summary, the BOS-TMC dataset represents a paradigm shift in computational transition metal chemistry by bridging the gap between experimental structural data and high-level electronic structure theory, while explicitly addressing the complexities of charge, spin, and methodological uncertainty.