Bond Dissociation Energy Skill

Goal

Calculate the homolytic and/or heterolytic bond dissociation energy (BDE) for each single bond in a molecule using Machine Learning Interatomic Potentials (MLIPs).

Homolytic BDE (radical fragments): $$\text{BDE}_\text{homo}(A{-}B) = E(A\bullet) + E(B\bullet) - E(A{-}B)$$

Heterolytic BDE (ionic fragments, minimum over both polarity variants): $$\text{BDE}_\text{hetero}(A{-}B) = \min!\bigl(E(A^+)+E(B^-),; E(A^-)+E(B^+)\bigr) - E(A{-}B)$$

[!IMPORTANT] This skill computes BDEs by relaxing both the intact molecule and fragments with an MLIP. For purpose-trained GNN models that predict BDE directly from SMILES (MAE ~0.6 kcal/mol), consider ALFABET or BonDNet instead.

Background

BDE is a fundamental thermodynamic quantity that determines:

• Drug metabolism: CYP450 enzymes abstract H from the weakest C–H bond
• Electrolyte stability: Which bonds break first under electrochemical voltage
• Combustion chemistry: Rate-determining bond-breaking steps in fuel oxidation
• Polymer degradation: Weakest links in polymer backbone chains

A 2024 study (Zubatyuk et al., JCTC) demonstrated that MACE potentials achieve BDE RMSE of 1.37 kcal/mol for aliphatic C–H bonds in drug-like molecules, outperforming semi-empirical methods and ALFABET for BDE ranking.

1. Prerequisites

• Conda Environment: mace-agent (includes RDKit, ASE, and MACE)
• Input: SMILES string or structure file (.sdf, .mol2)
• RDKit: Required for bond identification and molecular fragmentation

2. Choosing a Foundation Potential

Refer to the foundation-potentials skill for model selection.

[!IMPORTANT] Model requirements by cleavage mode:

Mode	Recommended model	supports_charge_spin	Validated?
homolytic	MACE-OFF23-small/medium/large	Not required	✅
heterolytic or both	MACE-OMOL-extra-large (env: mace-agent)	✅ Required	✅
heterolytic or both	MACE-MH-1 with omol head (env: mace-agent)	✅ Required	✅
heterolytic or both	FairChem uma-s-1p1 with --task_name omol (env: fairchem-agent)	✅ Required	✅

Setting charge/spin on MACE models: use atoms.info["charge"] and atoms.info["spin"] (the calculator's default info_keys maps "charge" → total_charge / "spin" → total_spin). Both MACE-OMOL and MACE-MH use joint_embedding to condition the network on these scalars.

If you request --cleavage both with a model that does not support charge/spin, the skill will log a warning and silently fall back to homolytic-only. Using --cleavage heterolytic with an unsupported model raises an error.

Note on single-atom fragments: When a bond produces a bare H (or other single atom), heterolytic BDE is automatically skipped — neither MACE nor FairChem UMA has signed single-atom energies (only neutral H, C, N, O… are in the reference tables).

3. Calculation Workflow

Step 1: Provide a molecule

# SMILES input (most common)
--smiles "CCO"

# Or from a structure file
--structure molecule.sdf

Step 2: Run BDE calculation

Homolytic only (default, no charge/spin needed):

# Env: mace-agent
python .agents/skills/chem-bond-dissociation/scripts/calculate_bde.py \
    --smiles CCO \
    --all_bonds \
    --cleavage homolytic \
    --model_type mace \
    --model_name MACE-OFF23-small \
    --output_dir research/my_folder/bde_results

Both homolytic and heterolytic (MACE-OMOL):

# Env: mace-agent
python .agents/skills/chem-bond-dissociation/scripts/calculate_bde.py \
    --smiles CCO \
    --all_bonds \
    --cleavage both \
    --model_type mace \
    --model_name MACE-OMOL-extra-large \
    --output_dir research/my_folder/bde_results_both

Both homolytic and heterolytic (FairChem UMA omol):

# Env: fairchem-agent
python .agents/skills/chem-bond-dissociation/scripts/calculate_bde.py \
    --smiles CCO \
    --all_bonds \
    --cleavage both \
    --model_type fairchem \
    --model_name uma-s-1p1 \
    --task_name omol \
    --output_dir research/my_folder/bde_results_both

Heterolytic only with FairChem UMA:

# Env: fairchem-agent
python .agents/skills/chem-bond-dissociation/scripts/calculate_bde.py \
    --smiles CCO \
    --all_bonds \
    --cleavage heterolytic \
    --model_type fairchem \
    --model_name uma-s-1p1 \
    --task_name omol \
    --output_dir research/my_folder/bde_hetero

Key Parameters

Argument	Default	Description
--smiles	—	SMILES string of the molecule
--structure	—	Path to structure file (.sdf, .mol2)
--bond	—	Specific bond as atom indices "i-j" (0-indexed)
--all_bonds	True	Compute BDE for all single bonds
--include_h_bonds	False	Include X–H bonds
--cleavage	homolytic	homolytic, heterolytic, or both
--model_type	mace	MLIP backend (mace, fairchem)
--model_name	auto	Model checkpoint (default: MACE-OFF23-small for homolytic; uma-s-1p1 for hetero/both)
--task_name	—	Task head for multi-task models (e.g. omol for FairChem UMA)
--fmax	0.01	Force convergence for relaxation (eV/Å)
--output_dir	required	Output directory

4. Output Files

• bde_results.json — Full results including:

  • metadata: model name, cleavage mode, supports_charge_spin, SMILES, etc.
  • intact_energy_eV: Energy of the relaxed intact molecule
  • bonds: List of per-bond results:
    • bde_eV, bde_kJ_mol, bde_kcal_mol: Homolytic BDE (if computed)
    • heterolytic_bde_eV, heterolytic_bde_kJ_mol, heterolytic_bde_kcal_mol: Best heterolytic BDE (if computed)
    • heterolytic_best_variant: Which polarity won ("frag1+ / frag2-" or "frag1- / frag2+")
    • heterolytic_variants: Raw results for both polarity variants
  • weakest_bond_homolytic, weakest_bond_heterolytic: Summary of weakest bonds
  • bonds_ranked_by_homolytic_bde, bonds_ranked_by_heterolytic_bde: Sorted tables

• intact_relaxed.xyz: Relaxed intact molecule

• frag_bond{N}_homo_{1,2}.xyz: Homolytic radical fragments

• frag_bond{N}_hetero_pos_neg_{1,2}.xyz: Heterolytic cation/anion fragments (variant A)

• frag_bond{N}_hetero_neg_pos_{1,2}.xyz: Heterolytic anion/cation fragments (variant B)

5. Examples

Example	Model	Cleavage	Notes
examples/ethanol_mace_off23_small/	MACE-OFF23-small	homolytic	Standard homolytic BDE for ethanol; includes H bonds
examples/methanol_mace_omol_both/	MACE-OMOL-extra-large	both	Homo + heterolytic for methanol; C–O hetero = 90 vs homo = 143 kcal/mol
examples/methanol_uma_omol_both/	FairChem UMA omol	both	Homo + heterolytic for methanol; C–O hetero = 157 vs homo = 126 kcal/mol

Ethanol — Homolytic BDE (MACE-OFF23)

# Env: mace-agent
python .agents/skills/chem-bond-dissociation/scripts/calculate_bde.py \
    --smiles CCO \
    --all_bonds \
    --include_h_bonds \
    --cleavage homolytic \
    --model_type mace \
    --model_name MACE-OFF23-small \
    --output_dir .agents/skills/chem-bond-dissociation/examples/ethanol_mace_off23_small

Methanol — Both Homo and Heterolytic BDE (FairChem UMA omol)

# Env: fairchem-agent
python .agents/skills/chem-bond-dissociation/scripts/calculate_bde.py \
    --smiles CO \
    --all_bonds \
    --include_h_bonds \
    --cleavage both \
    --model_type fairchem \
    --model_name uma-s-1p1 \
    --task_name omol \
    --output_dir .agents/skills/chem-bond-dissociation/examples/methanol_uma_omol_both

Experimental BDEs for ethanol (Blanksby & Ellison, 2003):

Bond	Experimental BDE (kcal/mol)
O–H	~104
C–H (methyl)	~101
C–H (methylene)	~95
C–C	~85
C–O	~92

6. Constraints

• Radical spin states: For homolytic BDE, MLIPs are generally "electron-agnostic" and treat fragments as neutral regardless of spin state. BDE ranking is typically more reliable than absolute values.
• Ionic states: Heterolytic BDE requires a charge/spin-aware model (supports_charge_spin=True). Validated: MACE-OMOL, MACE-MH (omol head), and FairChem UMA omol. These models use atoms.info["charge"] and atoms.info["spin"] to condition on the ionic state. Models without this flag raise an error for --cleavage heterolytic.
• Ring bonds: Breaking bonds in rings produces a single open-chain diradical. The script will warn and skip ring bonds.
• Accuracy: Expect ~2–5 kcal/mol error for homolytic BDEs with MACE-OFF23. Heterolytic accuracy is less benchmarked with current MLIPs.
• Environments:
  • mace-agent for MACE models
  • fairchem-agent for FairChem/UMA models

References

• Blanksby & Ellison, "Bond Dissociation Energies of Organic Molecules", Acc. Chem. Res. 2003, 36, 255.
• St. John et al., "Prediction of organic homolytic bond dissociation enthalpies at near chemical accuracy with sub-second computational cost", Nat. Commun. 2020, 11, 2328. (ALFABET)
• Zubatyuk et al., "A Transferable MACE Potential for Open- and Closed-Shell Drug-Like Molecules", J. Chem. Theory Comput. 2024.

---

Author: Bowen Deng Contact: GitHub @learningmatter-mit