Application: Inverse Design

Inverse design uses MolCraftDiffusion to generate molecules that satisfy a predefined objective rather than simply resembling the training data. Two common objective types are especially relevant:

Route	Objective	Mechanism
Property-directed	Target a scalar property (energy, excitation, score)	CFG, gradient guidance, or hybrid
Geometry-directed	Target a 3D arrangement (distance, angle, spatial motif)	Structure-guided generation + descriptor filtering

---

1. Target Property Inverse Design

Use this route when you want to bias generation toward a numerical target such as an energy score, excitation-related objective, or another learned molecular property.

Conceptual view: the diffusion model is combined with guidance models so generation is biased toward molecules with the desired target profile, followed by ranking and computational validation.

Concept

The model is no longer asked to generate molecules that are only plausible. It is asked to generate molecules that are plausible and biased toward a desired property target.

Note

This is the right framing when your question is:

• Can I shift generation toward a target electronic or physicochemical regime?

• Can I enrich the output in molecules with better scores?

• Can I search for hits without manually enumerating structures?

MolCraftDiffusion supports this through classifier-free guidance (CFG), gradient guidance (GG), and hybrid guidance.

Where to Go Next

See also

• Tutorial 7: Property-Directed Generation

• Tutorial 8: Property Prediction and Evaluation

• Tutorial 9: Analysis

Starting templates:

Config	Guidance mode
docs/cfg_examples/gen_cfg.yaml	Classifier-free guidance
docs/cfg_examples/gen_gradient_guidance.yaml	Gradient guidance
docs/cfg_examples/gen_hybrid_cfg_gg.yaml	Hybrid CFG + GG

Application scripts (singlet-fission use case):

• scripts/applications/singlet_fission/extract_gen_hit.py

• scripts/applications/singlet_fission/extract_g16_output.py

---

2. Target Geometry Inverse Design

Use this route when the design objective is primarily structural — for example a target distance, angle, or spatial arrangement between functional groups.

Conceptual view: a core motif is preserved while structure-guided generation builds candidates that satisfy a target 3D arrangement.

Concept

In this setting, the target is not just a scalar property. The target is a geometric arrangement that the generated molecule should realize in 3D space.

Note

This is the right framing when your question is:

• Can I generate molecules that place key atoms at the right distance or angle?

• Can I keep a reactive core while varying the surrounding structure?

• Can I enrich for candidates that satisfy a geometry-driven design hypothesis?

This typically combines structure-guided generation with descriptor-based filtering and downstream evaluation.

Where to Go Next

See also

• Tutorial 6: Structure-Guided Generation

• Tutorial 8: Property Prediction and Evaluation

• Tutorial 9: Analysis

Application scripts (IFLP geometry-targeting use case):

• scripts/applications/iflp_geom_target/compute_iflp_geom_desc.py

• scripts/applications/iflp_geom_target/process_int2_database.py

---

3. General Advice

Important

Keep these principles in mind for any inverse design campaign:

1. Start permissive, tighten later. Use relaxed generation settings first and apply stricter selection criteria once you see the output distribution.

2. Separate validity from objective satisfaction. High target scores alone are not enough — confirm the molecule is chemically reasonable.

3. Re-rank with external calculations. Physics- or chemistry-based checks (xTB, DFT) matter for any high-value campaign.

4. Treat guidance strength as a trade-off. Stronger guidance improves objective alignment but can reduce structural realism.