InvDesFlow: AI Inverse Design Workflow

• The paper presents InvDesFlow as an AI-enabled workflow that uses generative diffusion models and GNNs to design novel crystalline materials for superconductivity.
• It achieves high classification accuracy and low formation-energy prediction errors by integrating screening, first-principles validation, and an active-learning loop.
• InvDesFlow also features a tuning-free image-editing framework that preserves non-target content via authentic inversion and adaptive invariance control.

Searching arXiv for InvDesFlow and closely related papers to ground the article. InvDesFlow is a name used in recent arXiv literature for two distinct research systems. In one usage, it denotes an AI-driven inverse design workflow for crystalline materials, developed to discover new high-temperature conventional superconductors by combining generative crystal synthesis, superconductivity classification, stability prediction, superconducting-transition-temperature screening, and first-principles validation. In another usage, it denotes a tuning-free image-editing framework for rectified-flow text-to-image transformers, where the central problems are authentic inversion and invariance control. The materials-science usage is the more extensively elaborated one in the cited record, spanning general workflow papers, hydride-discovery studies, and an active-learning extension (Han et al., 2024, Ouyang et al., 21 Jan 2025, Han et al., 14 May 2025, Yao et al., 1 Aug 2025, Xu et al., 2024).

1. Terminological scope and research identity

In the materials literature, InvDesFlow is described as an “AI search engine” and, more broadly, as an “AI-driven inverse design workflow” for discovering high-$T_c$ superconductors. Its stated purpose is to move beyond searches confined to known databases by generating new crystal structures from scratch, filtering them for superconductivity likelihood and stability, and then validating selected candidates with physics-based calculations (Han et al., 2024).

In the image-editing literature, the same name is used for a tuning-free framework built on Stable Diffusion 3.5 / MM-DiT, a rectified-flow transformer. There the term does not refer to materials discovery at all; instead, it denotes a method for projecting a real image into the model’s domain and preserving non-target content during prompt-based editing (Xu et al., 2024).

Because these usages are separate, the term “InvDesFlow” is best understood as a shared project name rather than a single cross-domain methodology. In the materials-science branch, inverse design means starting from a desired property target—high superconducting critical temperature and structural stability—and working backward to generate candidate crystal structures. In the image-editing branch, inversion refers to reconstructing the flow-model trajectory of a real image, while invariance refers to preserving the parts of the image that should not change.

2. Materials inverse design workflow

The materials version of InvDesFlow is organized as a multi-stage pipeline. The general workflow reported in the superconductivity-discovery paper consists of symmetry-constrained crystal generation, superconductivity classification, formation-energy prediction, $T_c$ prediction with ALIGNN, physics-based validation, and active learning (Han et al., 2024). In the hydride papers, the same logic appears in a more targeted form: InvDesFlow performs the first-stage search over ternary or quaternary hydrides, proposes promising candidates, and then hands them off to density-functional theory, phonon calculations, electron-phonon coupling calculations, anisotropic Eliashberg calculations, and thermodynamic analysis (Ouyang et al., 21 Jan 2025, Yao et al., 1 Aug 2025).

Stage	Function	Reported role
Symmetry-constrained crystal generation	Generate new crystal structures	Search beyond existing databases
Superconductivity classification	Determine whether a generated structure resembles a superconductor	Filter candidates before expensive calculations
Formation-energy prediction	Estimate formation energy as a proxy for thermodynamic stability	Stability screening
$T_c$ prediction	Predict superconducting transition temperature	Prioritize high-$T_c$ candidates
Physics-based validation	Verify electronic structure, phonons, EPC, dynamical stability, and $T_c$	Final verification
Active learning	Fold validated discoveries back into training	Expand the search space iteratively

This pipeline is inverse-design-oriented because the design target is built in from the beginning. Rather than starting from a fixed catalog of known compounds, the workflow generates candidates, screens them by stability and superconducting potential, and only then applies first-principles validation. The hydride studies emphasize that InvDesFlow serves as the front end of the discovery pipeline, while DFT and EPC calculations serve as the verification and characterization stages (Ouyang et al., 21 Jan 2025).

3. Generative, discriminative, and screening components

A defining feature of InvDesFlow in the materials setting is the combination of generative and predictive models. In the general workflow paper, a diffusion model generates new crystal structures conditioned on atom count and symmetry constraints, and an E(n)-equivariant graph neural network serves as the denoiser. The crystal unit cell is represented by atom types $A$, fractional coordinates $X$, lattice matrix $L$, and atom number $N$, with the generative distribution written as

$p(C,N)=p(N)p(C\mid N).$

The paper also states that the wrapped normal distribution is used for periodic coordinates, and that sampling uses a predictor-corrector sampler followed by ASE and L-BFGS geometry optimization (Han et al., 2024).

The classification and screening side is equally explicit. For superconductivity classification, the workflow uses a graph auto-encoder / GNN framework with encoder, Wasserstein-based decoder, and pooling + softmax for classification. The pre-training data preparation uses 144,595 crystal entries from the Materials Project, while superconductivity fine-tuning uses 105 BCS superconductors with $T_c$0 K. The resulting classifier is reported to achieve a 99.04% discrimination success rate on those 105 BCS superconductors (Han et al., 2024).

For stability screening, the formation-energy predictor is trained on 380,000 crystal structures from GNoME and 60,000 from the Materials Project, giving 440,000 records in total. The paper reports that the improved model achieves 21 meV per atom formation-energy prediction error, compared with 28 meV for MEGNET, 39 meV for CGCNN, 35 meV for SchNet, 104 meV for CFID, and 930 meV for MAD (Han et al., 2024).

The quaternary-hydride paper describes the same architecture at a higher level as an integrated pipeline with two complementary parts: a generative model and a discriminant model. There, the crystal structure is reparameterized using a graph neural network representation, embedded into a high-dimensional latent space, and generated by a diffusion-based process that learns “reasonable element composition, atomic position, and crystal structure information.” The discriminative component rapidly outputs properties such as formation energy and superconducting temperature during the search (Yao et al., 1 Aug 2025). This suggests that InvDesFlow is not only a structure generator but also a property-aware ranking engine.

4. Superconductivity discoveries and hydride design logic

The first broad claim of the materials program is that InvDesFlow can generate new superconducting candidates outside existing databases. The general superconductivity paper reports 74 dynamically stable materials with critical temperatures predicted by the AI model to be $T_c$1 K, and states that these materials are not contained in any existing dataset. Three representative candidates were selected for DFT validation, of which $T_c$2 and $T_c$3 emerged as the two flagship cases (Han et al., 2024).

For $T_c$4, the workflow reports a metallic electronic structure, dominance of B, C, and N $T_c$5 orbitals near the Fermi level, an EPC constant $T_c$6, dynamical stability at ambient pressure, and a DFT/EPC transition temperature of

$T_c$7

For $T_c$8, the paper reports an imaginary phonon mode of about $T_c$9 meV along the $T_c$0 path at ambient pressure, disappearance of that instability at 5 GPa, $T_c$1, and

$T_c$2

under pressure (Han et al., 2024).

The most prominent hydride case is $T_c$3, found by InvDesFlow in a scan of ternary hydrides. The reported structural features are a cubic 216-type hydride, isostructural to $T_c$4, with Au-H octahedral motifs, and Li atoms intercalated into the interstitial sites between the octahedra. Li occupies Wyckoff site $T_c$5 at $T_c$6. Thermodynamic analysis proposes the ambient-pressure synthesis route

$T_c$7

with a calculated formation-energy difference of about

$T_c$8

which the paper states is well below the commonly used empirical metastability threshold of about

$T_c$9

DFT further shows no imaginary phonon modes, a metallic band structure, one band crossing the Fermi level, a density of states near $T_c$0 dominated by Au-H octahedral states, and a van Hove singularity at the $T_c$1 point. The superconducting result is $T_c$2 under ambient pressure, with a total EPC constant

$T_c$3

The key phonon findings are an $T_c$4 mode at $T_c$5 around 140 meV, an $T_c$6 mode at $T_c$7 around 20 meV, and an $T_c$8 mode at $T_c$9 around 30 meV. The paper emphasizes that the low-frequency modes involving Li contribute about 70% of the total $T_c$0 below 30 meV, and the anisotropic Eliashberg calculation gives a superconducting gap of about 26 meV at 55 K that vanishes near 140 K with $T_c$1 (Ouyang et al., 21 Jan 2025).

The later quaternary-hydride study extends this logic by atom intercalation. It searches $T_c$2 systems with $T_c$3 K, Na; $T_c$4 Cu, Ag; and $T_c$5 Ga, Li, all in the $T_c$6 space group. The central physical conclusion is that intercalating atoms could cause phonon softening and induce more phonon modes with strong electron-phonon coupling. In validated cases, the paper reports $T_c$7 K for $T_c$8, $T_c$9 K for $A$0, $A$1 K for $A$2, and $A$3 K for $A$4, compared with $A$5 K for the parent $A$6 (Yao et al., 1 Aug 2025).

Compound	Reported superconducting result	Context
$A$7	$A$8 K	Ambient pressure
$A$9	$X$0 K	5 GPa
$X$1	$X$2 K	Ambient pressure
$X$3	$X$4 K	Quaternary hydride
$X$5	$X$6 K	Quaternary hydride
$X$7	$X$8 K	Quaternary hydride

5. Active-learning extension: InvDesFlow-AL

InvDesFlow-AL is the active-learning extension of the materials workflow. It is described as a framework that pretrains a general crystal generator, fine-tunes it on a target functional-material dataset, generates candidate crystal structures, and then selects informative or valuable samples using a query-by-committee-style scoring rule before retraining the generator (Han et al., 14 May 2025).

The crystal representation remains a unit cell with atom types $X$9, fractional coordinates $L$0, and lattice matrix $L$1, and the generator is based on an EGNN. The total training loss is decomposed as

$L$2

The pretraining datasets are Alex-MP-20: 607,683 crystalline materials from MatterGen and GNoME: about 381,000 inorganic materials. The supplementary training details report 512 hidden dimensions, 6 GNN layers, 1000 diffusion steps, a 7.0 Å cutoff radius, 20 max neighbors, Adam with learning rate $L$3, and 1000 epochs on RTX 4090 (Han et al., 14 May 2025).

For crystal-structure prediction, the paper reports the headline result

$L$4

on MP-20, described as a 32.96% improvement over existing generative models. The comparison table gives 0.1045 Å for CDVAE, 0.0631 Å for DiffCSP, 0.0437 Å for CrystaLLM, 0.0510 Å for EquiCSP, and 0.0423 Å for InvDesFlow-AL (Han et al., 14 May 2025).

For low-formation-energy materials, the model is fine-tuned on GNoME structures with

$L$5

Over five iterations, the reported average formation energies become progressively more negative: $L$6 with generated counts of 80,707, 95,580, 97,379, 136,784, and 166,663, for a total of 577,113 generated crystals. For low-$L$7 discovery, the paper reports 1,598,551 materials with $L$8 meV after 10 fine-tuning rounds (Han et al., 14 May 2025).

The active-learning version also specializes the superconductivity pipeline. It introduces SuperconGNN, trained on 626 conventional superconductors from the Choudhary & Garrity dataset plus 59 hydride structures from the ambient-pressure hydride search literature, and uses a superconductivity query score that depends on $L$9, structural relaxation, novelty, and $N$0 meV. In this setting, the paper states that InvDesFlow-AL successfully identified $N$1 as a conventional BCS superconductor with a transition temperature of 140 K, while the active-learning loop improved the success rate of the DPA-2 atom-docking/post-processing model from 15% to 53% (Han et al., 14 May 2025, Han et al., 2024).

6. Image-editing usage: inversion and invariance in rectified-flow transformers

A separate paper uses the same name for a tuning-free image-editing framework for flow transformers. In that usage, InvDesFlow is built on Stable Diffusion 3.5 / MM-DiT, a rectified-flow transformer, and the central claim is that flow-based editing requires authentic inversion and flexible invariance control at the same time (Xu et al., 2024).

The rectified-flow formulation is written as

$N$2

with velocity field

$N$3

training objective

$N$4

and Euler sampling

$N$5

The paper’s inversion claim is that Euler inversion is structurally similar to DDIM inversion but more fragile because the reverse Euler step depends on $N$6, which is unknown during inversion. To address this, the method proposes a two-stage flow inversion. The first stage is fixed-point inversion,

$N$7

initialized with $N$8, and averaged after $N$9 iterations as

$p(C,N)=p(N)p(C\mid N).$0

The second stage is velocity compensation, with

$p(C,N)=p(N)p(C\mid N).$1

This residual is added during forward regeneration so that the reconstruction exactly follows the inversion trajectory (Xu et al., 2024).

Its invariance mechanism is not based on attention injection but on the text features inside adaptive layer normalization (AdaLN). Let $p(C,N)=p(N)p(C\mid N).$2 denote source and target text features in AdaLN before attention in an MM-DiT block. The method defines a token-aware mapping

$p(C,N)=p(N)p(C\mid N).$3

where $p(C,N)=p(N)p(C\mid N).$4 replaces the features of unedited target-prompt tokens with their source-prompt counterparts. This allows rigid and non-rigid editing while preserving non-target content. The paper explicitly mentions visual text changes, quantity changes, facial expressions, pose, layout, and object replacement. Its experiments are conducted on the PIE benchmark, containing 700 natural and artificial images, with baselines including Prompt-to-Prompt, Plug-and-Play, MasaCtrl, InfEdit, and InstructPix2Pix. The implementation uses CFG $p(C,N)=p(N)p(C\mid N).$5 for inversion, CFG $p(C,N)=p(N)p(C\mid N).$6 for editing, 30 inversion steps, and a fixed-point iteration number of 3 (Xu et al., 2024).

7. Significance, assumptions, and limitations

Across the materials papers, InvDesFlow matters because it is presented as a way to search a very large crystal-composition space more efficiently than trial-and-error or brute-force structure search. The general workflow paper argues that the materials universe is far larger than existing datasets, and the reported 74 novel, dynamically stable candidates with predicted $p(C,N)=p(N)p(C\mid N).$7 K are offered as evidence that the method can generate crystal structures not contained in current databases (Han et al., 2024). The hydride studies strengthen that claim by showing that the AI search engine can target a chemically restricted space—ambient-stable superconducting hydrides—and still identify candidates with a credible synthesis route, dynamical stability, and strong EPC signatures (Ouyang et al., 21 Jan 2025, Yao et al., 1 Aug 2025).

At the same time, the materials papers state several constraints. The general workflow validates only three representative candidates with DFT rather than all 74. The quaternary-hydride paper notes that the study remains template-based, focusing on atom-intercalated structures derived from 216-type ternary hydrides, and that the exact formation-energy cutoff is not given in the main text. The active-learning paper further notes dependence on pretraining data quality, continued reliance on DFT in the query-by-committee loop, and sensitivity to the design of the scoring rules (Han et al., 2024, Yao et al., 1 Aug 2025, Han et al., 14 May 2025).

The image-editing paper states a different limitation profile. Its framework depends on the real image being reasonably representable by the flow model’s prior; if the image is significantly out-of-domain, the inversion trajectory may deviate too much from the authentic generation process. It also notes computational overhead, since fixed-point inversion requires multiple transformer evaluations per timestep (Xu et al., 2024).

Taken together, these papers define InvDesFlow less as a single fixed algorithm than as a research program organized around inverse design and flow-based control. In materials science, the name denotes an AI search engine that couples generative modeling with property prediction and first-principles verification; in flow-transformer image editing, it denotes a tuning-free framework for inversion and invariance. The common thread is the use of learned generative priors as the front end of a targeted search process, but the technical content and application domains are otherwise distinct.