Periodic Materials Generation using Text-Guided Joint Diffusion Model

TGDMat(Short) does not always outperform baseline methods.

In Table 3, the text-guided baseline models such as CDVAE+, SyMat+, and DiffCSP+ are based on the long variant of their respective models. As a result, TGDMat(Short) does not consistently outperform these baseline methods, although it does outperform them in 7 out of 21 cases, ranking as the second-best model overall. To provide a more equitable comparison, we have separated the results of the short variants in the following table for both tasks.

Gen Task (on Shorter Prompt) :

Perov Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
CDVAE+(Short)	98.17	100	99.40	99.01	0.0706	0.1395	0.0246
SyMat+(Short)	96.94	100	99.22	98.40	0.0192	0.1827	0.2633
DiffCSP+(Short)	98.21	100	99.61	98.39	0.0123	0.1193	0.0266
TGDMat(Short)	98.28	100	99.71	99.24	0.0108	0.0947	0.0237

Carbon Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
CDVAE+(Short)		100	99.34	82.96		0.1398	0.2804
SyMat+(Short)		100	99.52	97.20		0.1206	3.7422
DiffCSP+(Short)		100	99.65	97.29		0.0811	0.0870
TGDMat(Short)		100	99.81	91.77		0.0681	0.0865

MP Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
CDVAE+(Short)	87.05	100	99.36	99.60	0.9930	0.6420	0.2970
SyMat+(Short)	88.08	99.9	98.59	99.47	0.5031	0.3917	0.3622
DiffCSP+(Short)	84.57	100	99.52	99.85	0.3310	0.3800	0.1379
TGDMat(Short)	86.60	100	99.79	99.88	0.3337	0.3296	0.1189

We observe that among the total 19 metrics over three datasets in the table, TGDMat(Short) method outperforms all other baselines in 16 metrics.

CSP Task (on Shorter Prompt) :

Method	#Samples	Perov	Perov	Carbon	Carbon	MP	MP
		Match Rate	RMSE	Match Rate	RMSE	Match Rate	RMSE
CDVAE+(Short)	1	48.97	0.1063	22.65	0.2640	40.33	0.1037
CDVAE+(Short)	20	89.54	0.0423	89.61	0.2188	70.22	0.0876
SyMat+(Short)	1	49.39	0.0985	23.71	0.2567	40.84	0.1027
SyMat+(Short)	20	92.10	0.0255	90.86	0.2069	71.31	0.0875
DiffCSP+(Short) or TGDMat (Short)	1	56.54	0.0583	24.13	0.2424	52.22	0.0597
DiffCSP+(Short) or TGDMat (Short)	20	98.25	0.0137	88.28	0.2252	80.97	0.0443

In CSP Task, we observe TGDMat (Short) performs better than all the baselines across dataset in both setup: using number of samples 1 and 20. Please note in the CSP task, since atom types are given DiffCSP+ (Short) and TGDMat (Short) are the same architecture.

Marginal Improvement of TGDMat(Long) over baseline methods.

We respectfully disagree with the reviewer’s assessment that TGDMat(Long) offers only marginal improvements.

Gen Task (on Long Prompt) :

For metrics where baselines already achieve over 98% accuracy (e.g., Structural and Compositional Validity for Perov, Carbon, or Coverage Metrics), the scope for further improvement is inherently limited. However, for other metrics, TGDMat(Long) demonstrates significant enhancements. Specifically, the average improvement in property statistics are as follows:

	Perov	Carbon	MP
Avg Improvement by TGDMat(Long)	35.61%	38.91%	14.57%

Moreover, for the MP-20 dataset, compositional validity is improved by 5%, which is a notable advancement. Overall, TGDMat’s superior performance across seven metrics on the MP-20 dataset underscores its potential to generate novel materials that are experimentally synthesizable.

CSP Task (on Long Prompt) :

Prior unconditional diffusion models demonstrate improved Match Rates and lower RMSE when generating 20 samples (k = 20) per test material, they largely fail in both metrics when generating only one sample (k = 1) per test material. Using text guidance(Long Variant) during the reverse denoising process, with just one generated sample per test material, text-guided variants outperform respective vanilla models, thereby reducing computational overhead.

Overall, these findings highlight the impact of our key contributions in TGDMat: 1) Joint Diffusion on A, X, and L, 2) Discrete Diffusion on A, and 3) Text-Guided Diffusion.

Finally, we would like to thank Reviewer mJeo once again for these valuable comments. We will reflect these comments in the revised manuscript. We believe that our responses above address all of Reviewer mJeo's concerns and contribute to further strengthening our work.

Sincerely, The Authors

Dear Reviewer mJeo,

Thank you very much for providing us with valuable feedback. We appreciate the detailed comments. Below, we have provided point-by-point responses to each of your comments.

Weaknesses

• Regarding Architectural Innovation of TGDMat.

While text-guided diffusion models are widely used in fields such as image, video, and molecule generation, their application in periodic material generation remains largely unexplored. This study is the first to introduce text-guided diffusion for material generation. To achieve this, we have made the following novel contributions:

• Dataset Curation: At the start of this work, no textual data was available for materials in benchmark databases such as Perov, Carbon, and MP. To address this, we curated textual data for these material databases, including both long and short textual prompts. Details of this curation process are provided in Section 4.2 and Appendix D: Textual Dataset. We plan to release these datasets to the community, anticipating that they will facilitate further exploration and research.

• Simple Yet Effective Fusion Approach: In TGDMat, we employ a straightforward yet highly effective method to integrate text-based contextual representations into the denoising network. Specifically, at each stage of denoising, we utilize the contextual information $**C**_**p**$ derived from textual descriptions via a pre-trained MatSciBERT model and concatenate it with the input atom features.

Our innovation, however, goes beyond text-guided diffusion. We further enhance the base diffusion model by introducing Discrete Diffusion for Atom Types. Prior works like CDVAE, DiffCSP used continuous diffusion on atom types (represented as the probability distribution for k classes and applying DDPM to learn the distribution). However, atom types are discrete data and it is well established in the literature[1] that using a continuous diffusion model for discrete features is not reasonable and produces suboptimal results. Hence we consider discrete diffusion on atom types in TGDMat, where we consider A as N discrete variables belonging to k classes and leverage the discrete diffusion model (D3PM)

[1] Andrew Campbell et.al. A continuous time framework for discrete denoising models. Advances in Neural Information Processing Systems 35 (2022), 28266–28279.

• Regarding Diversity and Novelty as additional metrics for comparison.

• Regarding Novelty: For evaluation, we adopted the benchmark metrics introduced by CDVAE and subsequently used by prior works such as SyMat and DiffCSP. As outlined by Xie et al. in the CDVAE paper (Section 5.2: "Material Generation"), these seven metrics in the Generation Task aim to assess the novelty, validity, and property statistics of the generated materials.

• Regarding Diversity: Since our paper focuses on conditional material generation, the goal is not to achieve diverse random generation but rather to produce constrained and targeted outputs that align with the criteria specified in the text descriptions of the test dataset. Consequently, the diversity of the generated materials depends on the diversity of the test dataset—if the test dataset is diverse, the generated materials will reflect that diversity; otherwise, they will not.

Ablation study for the separately and jointly learning of crystal geometry.

In response to the reviewer's feedback we conducted an ablation study where we use three diffusion models to learn A,X,L separately. While sampling we sample A,X,L separately and merge them together. We fuse the textual representation in the same way in all three diffusion models. We present the results in following table and compare with TDGMat:

Perov Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
TGDMat (Learning A,X,L seperately)	90.10	85.43	85.77	83.51	0.341	0.591	0.376
TGDMat (Jontly Learning A,X,L)	98.63	100	99.87	99.52	0.0090	0.049	0.018

Carbon Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
TGDMat (Learning A,X,L seperately)		75.64	80.95	82.29		0.435	0.584
TGDMat (Jontly Learning A,X,L)		100	99.91	92.43		0.043	0.063

MP Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
TGDMat (Learning A,X,L seperately)	73.18	77.01	82.99	72.41	0.861	0.887	0.634
TGDMat (Jontly Learning A,X,L)	92.97	100	99.89	99.95	0.289	0.308	0.115

We observe significant performance degradation in all metrics across datasets if we learn A,X,L separately. We will add these results in the revised manuscripts.

What’s the difference between SE(3) and periodic E(3)-equivariant denoising model backbone?

A model is An SE(3) equivariant if it follows invariance to permutation, translation, and rotation. Additionally, for crystals, we need invariance for periodic transformations, since the atoms in the unit cell can periodically repeat themselves infinite times along the lattice vector, there can be many choices of unit cells and coordinate matrices representing the same material. We denote this as a periodic E(3)-equivariant denoising model or GNN. SE(3) equivariant GNNs adapted with periodicity proposed by CDVAE are also similar to the GNN model that follows invariance to permutation, translation, rotation, and periodicity. More details about the physical symmetry of the crystal structure are provided in Appendix C: "INVARIANCES IN CRYSTAL STRUCTURE."

How would one select the best sample if the ground truth was not available?

In the absence of ground truth, selecting the best sampleamong generated materials involves ensuring things such as the generated samples adhere to fundamental physical and chemical principles, such as proper bonding patterns, reasonable interatomic distances, and compliance with crystal symmetry constraints. One can also use tools like DFT to calculate formation energy, energy above the convex hull, or phonon stability to select the best samples. However, at this point of time there is no single agreed method to choose the best sample.

Finally, we would like to thank Reviewer ijYE once again for these valuable comments. We will reflect these comments in the revised manuscript. We believe that our responses above address all of Reviewer ijYE's concerns and contribute to further strengthening our work.

Sincerely, The Authors

Evaluation of whether the generated materials meet the text descriptions.

Quantitative results on the alignment of the generated structures with the given prompts, in terms of compositions, properties, and other factors, are provided in the "F.3 CORRECTNESS OF GENERATED MATERIALS" section of the Appendix (Page 23, Line 1229). This is also referenced in the "Additional Results" section (Page 10, Line 527) of the main manuscript. Following are the details :

To ensure the fidelity of our model’s outputs concerning these specified global attributes from the text prompt, we randomly generated 1000 materials (sampled from all three Datasets) based on their respective textual descriptions(both Long and Short) and assessed the percentage of generated materials that matched the global features outlined in the text prompt. In specific, we matched the Formula, Space group, Crystal System, and Dimensions of generated materials with the textual descriptions. Moreover, we examined whether properties such as formation energy and bandgap matched the specified criteria as per the text prompt (positive/negative, zero/nonzero).

Results for TGDMat(Long):

Global Feature	% of Matched Materials(Perov)	% of Matched Materials (Carbon)	% of Matched Materials(MP)
Formula	97.50	98.20	70.54
Space Group	87.00	80.79	67.88
Crystal System	92.60	91.55	73.54
Formation Energy	95.49	-	92.88
Band Gap	-	98.61	96.73

Results for TGDMat(Short):

Global Feature	% of Matched Materials(Perov)	% of Matched Materials (Carbon)	% of Matched Materials(MP)
Formula	90.70	92.56	65.22
Space Group	86.51	80.50	58.77
Crystal System	83.19	81.64	72.77
Formation Energy	90.33	-	91.00
Band Gap	-	95.90	93.33

In general, using longer text, considering Perov-5 and Carbon-24 datasets, the generated material meets the specified criteria effectively. However, when dealing with the MP-20 dataset, which is more intricate due to its complex structure and composition, performance tends to decline. Additionally, when using shorter prompts, overall performance suffers across all datasets compared to longer text inputs. This is because the longer text, provided by the robocrystallographer, offers a comprehensive range of information, both global and local, thereby enhancing the generation capabilities of TGDMat.

Questions

To address issues in local message passing, how about using Transformers?

Replacing GNNs with Transformers as the backbone network is an exciting research direction, but implementing Transformers comes up with its own challenges[1]. One of the key obstacles is the scalability and efficiency of graph transformers, which require significant memory and computational resources, particularly when using global attention mechanisms. These challenges become even more pronounced in deeper architectures, which are more prone to overfitting and over-smoothing. Additionally, graph transformers often face challenges when it comes to generalizing to graphs. As a result, in many popular graph machine learning applications, graph transformers have not yet fully replaced the Message Passing (GNN) framework. We consider this an important area for future exploration.

However, in response to the reviewer's suggestion we conduct an additional study, where we replace the backbone of TGDMat with Matformer[2], a popular transformer model for crystal property prediction. We present the results of Gen Tasks in the following table and compare them with TDGMat:

Perov Dataset:

Method	Comp Validity	Struct Validity
TGDMat (Matformer Backbone)	93.77	90.13
TGDMat (GNN Backbone)	98.63	100

Carbon Dataset:

Method	Comp Validity	Struct Validity
TGDMat(Matformer Backbone)	-	89.26
TGDMat(GNN Backbone)	-	100

MP Dataset:

Method	Comp Validity	Struct Validity
TGDMat(Matformer Backbone)	81.96	84.37
TGDMat(GNN Backbone)	92.97	100

We observe significant performance degradation in all metrics across datasets. This needs further exploration and we keep this a scope of future work.

[1] Shehzad, A. Graph transformers: A survey. arXiv preprint arXiv:2407.09777.

[2] Yan, Keqiang, et al." Periodic graph transformers for crystal material property prediction. NeuRIPs 35 (2022): 15066-15080.

Regarding Typo in Fig-2.

We apologize for the mistake and appreciate you pointing it out. The direction of the arrows for the diffusion and reverse processes was incorrect. For the diffusion/forward process, the arrow should point from $M_0$ to $M_T$, while for the denoising/reverse process, it should point from $M_T$ to $M_0$. Also, the GNN model would be a periodic E(3)-equivariant GNN model. We will correct this error and update the revised manuscript accordingly.

Dear Reviewer ijYE,

Thank you very much for providing us with valuable feedback. We appreciate the detailed comments. Below, we have provided point-by-point responses to each of your comments.

Weaknesses:

Utility of Text-guidace than Feature Vectors-guidance

In literature [1], it has been extensively studied that using value guidance specifying a single or a handful of target properties or features as feature vectors might be insufficient to capture intricate conditions. In contrast, textual descriptions allow us to encompass these conditions adeptly and flexibly and produce better results than value-based guidance. Also, Since we use MatSciBert which is pre-trained on a huge corpus of articles from materials domains, it allows us to encode the more enriched and robust contextual information about the tokens.

[1] Luo, Yanchen, et al. "Text-guided diffusion model for 3d molecule generation." (2024).

However, in response to the reviewer's question, we conducted an additional experiment, where we feed all relevant conditional information e.g Formula, Space Group, Crystal Symmetry,Bond Length and Property Values as feature vectors to guide the diffusion model. Here are the results we observed for Gen Task:

Perov Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
TGDMat (Conditions as Features)	96.55	98.73	99.18	97.06	0.0149	0.1200	0.0290
TGDMat (Conditions as Text Emb)	98.63	100	99.87	99.52	0.0090	0.049	0.018

Carbon Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
TGDMat (Conditions as Features)		99.56	99.32	92.17		0.104	0.087
TGDMat (Conditions as Text Emb)		100	99.91	92.43		0.043	0.063

MP Dataset:

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
TGDMat (Conditions as Features)	83.85	99.61	98.97	98.3	0.377	0.375	0.126
TGDMat (Conditions as Text Emb)	92.97	100	99.89	99.95	0.289	0.308	0.115

Across three datasets, we did not observe performance improvements. We will add these results in the revised manuscripts.

Key differences between DiffCSP and TGDMat

Following are the key differences between DiffCSP and our proposed TGDMat

	DiffCSP	TGDMat
Tasks	Only CSP Task	Both CSP and Gen Tasks
Diffusion on Atom Type	NA	Discrete Diffusion (D3PM)
Model Category	Unconditional; unable to specify the criteria required by the user.	Conditional; able to specify the criteria required by the user.(in Text Format)
Text Guided Diffusion	NO	Yes

However, note the goal of this paper is not to introduce a new diffusion model to replace existing models like DiffCSP or CDVAE for periodic material generation. Instead, we focus on demonstrating that conditional models can outperform traditional unconditional models, such as DiffCSP. Specifically, we show that incorporating textual conditions through text-guided diffusion leads to better performance compared to using unconditional models like DiffCSP. Additionally, we enhance DiffCSP by integrating discrete diffusion over atom types in our proposed TGDMat framework.

Regarding Diffusion Method Design

In the diffusion model, the forward process is non-parametric, meaning we simply add noise to three variables (atom coordinates, atom types, and the lattice) independently, without any learning involved. However, during the denoising process, a backbone Equivariant Graph Neural Network (EGNN) is used to predict the noise at each time step. At any given time t, the EGNN takes the atom types A_t, atom coordinates X_t, and lattice L_t together as input. Then the EGNN performs message passing and aggregation to generate node and graph representations, which are then used to predict the noise. Thus, the denoising process depends on all three variables.

I thank the authors for delivering additional results and discussions quickly to answer my questions. Incorporating the results and discussions can improve the soundness of the work. I increase the soundness score from 2 to 3. My evaluation on the contribution remains. I may put the overall score between 6 and 7, however, based on the assessment on novelty and contribution, it cannot reach 8 in my evaluation. So I'd keep the overall score the same.

Ablation study comparing long and short text prompts on the baseline models.

We begin by presenting the ablation study results comparing long and short text prompts on both tasks, followed by key insights and observations.

Gen Task:

Perov

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
CDVAE	98.29	100	99.25	98.39	0.0731	0.1462	0.0291
CDVAE+(Short)	98.37	100	99.40	99.01	0.0706	0.1395	0.0246
CDVAE+(Long)	98.45	100	99.53	99.09	0.0609	0.1276	0.0223
SyMat	96.83	100	99.16	98.29	0.0193	0.1991	0.2827
SyMat+(Short)	96.94	100	99.22	98.40	0.0192	0.1827	0.2633
SyMat+(Long)	97.88	100	99.71	98.79	0.0172	0.1755	0.2566
DiffCSP	98.15	100	99.28	98.08	0.0132	0.1281	0.0267
DiffCSP+(Short)	98.21	100	99.61	98.39	0.0123	0.1193	0.0266
DiffCSP+(Long)	98.44	100	99.85	98.53	0.0119	0.1071	0.0241
TGDMat(Short)	98.28	100	99.71	99.24	0.0108	0.094	0.023
TGDMat(Long)	98.63	100	99.87	99.52	0.009	0.049	0.018

Carbon

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
CDVAE		100	99.35	82.66		0.1539	0.2889
CDVAE+(Short)		100	99.34	82.96		0.1398	0.2804
CDVAE+(Long)		100	99.82	84.76		0.1377	0.2660
SyMat		100	99.42	97.17		0.1234	3.9628
SyMat+(Short)		100	99.52	97.20		0.1206	3.7422
SyMat+(Long)		100	99.90	97.63		0.1171	3.8620
DiffCSP		99.9	99.49	97.27		0.0861	0.0876
DiffCSP+(Short)		100	99.61	97.29		0.0811	0.087
DiffCSP+(Long)		100	99.93	97.33		0.0763	0.0853
TGDMat(Short)		100	99.81	91.77		0.0681	0.0865
TGDMat(Long)		100	99.9	92.43		0.043	0.063

MP

Method	Comp Validity	Struct Validity	Cov-R	Cov-P	#Element	Density	Energy
CDVAE	86.30	100	99.15	99.49	1.4921	0.7085	0.3039
CDVAE+(Short)	87.05	100	99.36	99.60	0.9930	0.6420	0.2970
CDVAE+(Long)	87.42	100	99.57	99.81	0.9720	0.6388	0.2977
SyMat	87.96	99.9	98.30	99.37	0.5236	0.4012	0.3877
SyMat+(Short)	96.94	99.9	98.59	99.47	0.5031	0.3917	0.3622
SyMat+(Long)	97.88	99.9	99.01	99.95	0.4865	0.3879	0.3489
DiffCSP	83.25	100	99.41	99.76	0.3411	0.3802	0.1497
DiffCSP+(Short)	84.57	100	99.52	99.85	0.3310	0.3800	0.1379
DiffCSP+(Long)	85.07	100	99.81	99.89	0.3122	0.3799	0.1355
TGDMat(Short)	86.60	100	99.79	99.88	0.3337	0.3296	0.1189
TGDMat(Long)	92.97	100	99.89	99.95	0.289	0.308	0.115

CSP Task :

Method	#Samples	Perov	Perov	Carbon	Carbon	MP	MP
		Match Rate	RMSE	Match Rate	RMSE	Match Rate	RMSE
CDVAE	1	45.31	0.1138	17.09	0.2969	33.90	0.1045
CDVAE	20	88.51	0.0464	88.37	0.2286	66.95	0.1026
CDVAE+(Short)	1	48.97	0.1063	22.65	0.2640	40.33	0.1037
CDVAE+(Short)	20	89.54	0.0423	89.61	0.2188	70.22	0.0876
CDVAE+(Long)	1	49.25	0.1055	23.73	0.2590	41.80	0.1021
CDVAE+(Long)	20	89.73	0.0417	89.77	0.2053	72.56	0.0840
SyMat	1	47.32	0.1074	20.81	0.2655	33.92	0.1039
SyMat	20	90.25	0.0316	89.29	0.2184	71.03	0.0945
SyMat+(Short)	1	49.39	0.0985	27.71	0.2567	40.84	0.1027
SyMat+(Short)	20	92.10	0.0255	90.86	0.2069	71.31	0.0875
SyMat+(Long)	1	50.88	0.0963	28.18	0.2510	43.17	0.1016
SyMat+(Long)	20	92.30	0.0201	91.65	0.1870	72.96	0.0820
DiffCSP	1	52.02	0.0760	17.54	0.2759	51.49	0.0631
DiffCSP	20	98.60	0.0128	88.47	0.2192	77.93	0.0492
DiffCSP+(Short)	1	56.54	0.0583	24.13	0.2424	52.22	0.0597
DiffCSP+(Short)	20	98.25	0.0137	88.28	0.2252	80.97	0.0443
DiffCSP+(Long)	1	90.46	0.020	44.63	0.226	55.15	0.057
DiffCSP+(Long)	20	98.59	0.007	95.27	0.153	82.02	0.039

Observations:

1. For both tasks, across all the datasets, text guidance out performs the vanilla diffusion models in almost all metrics.
2. Our experiments suggest that using shorter prompts textguided models outperforms the vanilla baseline models. However performance is even superior when using text guided diffusions using longer prompts.
3. For CSP task, using text guidance during the reverse denoising process, with just one generated sample per test material, text-guided variants outperform respective vanilla models, thereby reducing computational overhead.
4. Our proposed TGDMat (Long) stands out as the leading model when compared to all baseline models and their text-guided variants across three benchmark datasets. In specific, for Gen Task, TGDMat (Long) outperforms closest baseline DiffCSP+ (Long) because we leveraged discrete diffusion on atom types, which is more powerful learning discrete variables like atom types.
5. Finally, results indicate that utilizing shorter prompts TGDMat (Short) results in a slight decrease in overall performance compared to the longer variant TGDMat (Long). Nonetheless, the performance remains superior or comparable to baseline models (vanilla and text-guided variants)

We will add these comprehensive results in the revised manuscripts (Appendix F.3 in the revised version).

Dear Reviewer D75u,

Thank you very much for providing us with valuable feedback. We appreciate the detailed comments. Below, we have provided point-by-point responses to each of your comments.

Questions:

Details on how the contextual representation of long, detailed text data is integrated into the denoising network?

As mentioned in Section 4.3.2 TEXT GUIDED DENOISING NETWORK (Line-313-18), Equation-5, At each timestep t of reverse diffusion, we concatenate textual representation $**C**_**p**$ with each input atom feature. Following this same approach, we also developed text-guided versions of the baseline models, named CDVAE+, SyMat+, and DiffCSP+, in which the contextual representation from detailed text data is integrated into the denoising networks of these models.

How well generated materials align with the given text prompts?

Quantitative results on the alignment of the generated structures with the given prompts, in terms of compositions, properties, and other factors, are provided in the "F.3 CORRECTNESS OF GENERATED MATERIALS" section of the Appendix (Page 23, Line 1229). This is also referenced in the "Additional Results" section (Page 10, Line 527) of the main manuscript. In the revised manuscript, this has been incorporated into Section 5.4 of the main manuscript.

Following are the details :

To ensure the fidelity of our model’s outputs concerning these specified global attributes from the text prompt, we randomly generated 1000 materials (sampled from all three Datasets) based on their respective textual descriptions(both Long and Short) and assessed the percentage of generated materials that matched the global features outlined in the text prompt. In specific, we matched the Formula, Space group, and Crystal System, and Dimensions of generated materials with the textual descriptions.Moreover, we examined whether properties such as formation energy and bandgap matched the specified criteria as per the text prompt (positive/negative, zero/nonzero).

Results for TGDMat(Long):

Global Feature	% of Matched Materials(Perov)	% of Matched Materials (Carbon)	% of Matched Materials(MP)
Formula	97.50	98.20	70.54
Space Group	87.00	80.79	67.88
Crystal System	92.60	91.55	73.54
Formation Energy	95.49	-	92.88
Band Gap	-	98.61	96.73

Results for TGDMat(Short):

Global Feature	% of Matched Materials(Perov)	% of Matched Materials (Carbon)	% of Matched Materials(MP)
Formula	90.70	92.56	65.22
Space Group	86.51	80.50	58.77
Crystal System	83.19	81.64	72.77
Formation Energy	90.33	-	91.00
Band Gap	-	95.90	93.33

In general, using longer text, considering Perov-5 and Carbon-24 datasets, the generated material meets the specified criteria effectively. However, when dealing with the MP-20 dataset, which is more intricate due to its complex structure and composition, performance tends to decline. Additionally, when using shorter prompts, overall performance suffers across all datasets compared to longer text inputs. This is because the longer text, provided by the robocrystallographer, offers a comprehensive range of information, both global and local, thereby enhancing the generation capabilities of TGDMat.

Dear Reviewer,

Thank you for your valuable feedback and constructive comments.

In our rebuttal, we have provided additional experiments and enhanced explanations, and we hope we have addressed all the concerns raised by the reviewer. We are open to further discussions and are happy to clarify any remaining doubts.

If there are still any outstanding issues, we kindly request you to share them with us. Otherwise, we would greatly appreciate it if you could consider revising the score.

We look forward to your response.

Thank you,

The Authors

Dear Reviewer cwRr,

Thank you very much for providing us with valuable feedback. We appreciate the detailed comments. Below, we have provided point-by-point responses to each of your comments.

Weaknesses:

Key differences between DiffCSP and TGDMat

Following are the key differences between DiffCSP and our proposed TGDMat

	DiffCSP	TGDMat
Tasks	Only CSP Task	Both CSP and Gen Tasks
Diffusion on Atom Type	NA	Discrete Diffusion (D3PM)
Model Category	Unconditional; unable to specify the criteria required by the user.	Conditional; able to specify the criteria required by the user. (in Text Format)
Text Guided Diffusion	NO	Yes

However, note the goal of this paper is not to introduce a new diffusion model to replace existing models like DiffCSP or CDVAE for periodic material generation. Instead, we focus on demonstrating that conditional models can outperform traditional unconditional models, such as DiffCSP. Specifically, we show that incorporating textual conditions through text-guided diffusion leads to better performance compared to using unconditional models like DiffCSP. Additionally, we enhance DiffCSP by integrating discrete diffusion over atom types in our proposed TGDMat framework.

Suggestion: Use of \citep command.

Thanks for the suggestion. We will update the revised manuscript.

Regarding Practical Deployment of the Model in Industry.

Our proposed method demonstrates significant potential for practical material discovery in industry by generating valid, diverse, and structurally plausible materials aligned with user-provided textual descriptions. This capability positions our model as an efficient tool for creating “initial templates” of materials tailored for applications such as battery materials, solar cells, or catalysts, significantly reducing the time and computational resources required for exploratory studies.

However, some barriers must be addressed before full deployment in industrial workflows. One challenge is the potential mismatch between generated structures and experimental ground truth, which arises from the inherent approximations in the model. These generated structures require further validation and refinement using computational methods, such as density functional theory (DFT), to ensure their physical and chemical feasibility before full deployment in industrial workflows.

Questions:

Model Size of all Baselines and TGDMat

Models	# Parameters	Model size
CDVAE	4920414	18.771 MB
SyMat	3385601	12.915 MB
DiffCSP	12294656	46.923 MB
TGDMat	12432228	47.448 MB

We would like to once again express our gratitude to Reviewer cwRr for their valuable comments and suggestions. We will incorporate these insights into the revised manuscript. We believe our responses above effectively address all of Reviewer cwRr's concerns and further enhance the quality of our work.

Sincerely,
The Authors

Dear Reviewer,

Thank you for your valuable feedback and constructive comments.

In our rebuttal, we have provided additional experiments and enhanced explanations, and we hope we have addressed all the concerns raised by the reviewer. We are open to further discussions and are happy to clarify any remaining doubts.

If there are still any outstanding issues, we kindly request you to share them with us. Otherwise, we would greatly appreciate it if you could consider revising the score.

We look forward to your response.

Thank you,

The Authors