Extending Galactic Foreground Emission with Neural Networks Journal: Astronomy & \& Computing Giuseppe Puglisi, Avinash Anand, Marina Migliaccio Abstract We introduce an innovative approach employing Cycle Generative Adversarial Networks (Cycle-GANs) to accurately simulate Carbon Monoxide (CO) emissions by learning features identified in thermal dust emission maps from the Planck satellite alongside HI data from HI4PI survey. Our training dataset is complemented by the targets represented by the two rotational transition lines of CO ( J : 1 − 0 , 2 − 1 J:1-0,\,2-1 ) provided by the Planck satellite. We ensure the robustness of our dataset by focusing on regions with a signal-to-noise ratio (SNR) exceeding 8. The outcomes, assessed utilizing angular power spectra and Minkowski functionals, confirm that our algorithm proficiently achieves the set goals, indicating that the amplitudes of the generated emission accurately reproduce the angular correlations and share the statistical properties of the employed CO targets. We thus aim at improving the current models of CO emission specifically in the high-Galactic latitude areas that have been hardly observed by the most recent surveys, and, in doing so, to address and overcome the limitations affecting current models regions. This research lays the groundwork for creating transformative synthetic simulations, leveraging convolutional neural networks tied to data procured from latest observations. 1 Introduction The emissions of Galactic foregrounds significantly contaminate the Cosmic Microwave Background (CMB) measurements. To address this challenge, recent efforts in the scientific community have focused on acquiring multi- wavelength observations, spanning 0.1 to 1 mm wavelength. This multi-frequency approach enables accurate characterization of the spatial and spectral distributions of each foreground component, crucial for isolating the cosmological signal. The sub-mm emission of our own Galaxy includes: thermal dust, synchrotron and molecular line emissions ( 20 ) . Dust grains heated by optical starlight emit thermal radiation in the Far Infra-red and microwave regime ( \nu>90 GHz); Galactic synchrotron radiation, due to cosmic electrons spiraling into the magnetic field, mainly contributes at low frequencies ( \nu<70 GHz). Both synchrotron and dust show large levels of polarization ( up to 18%, 20 ). The Carbon Monoxide (CO) emission in the sub-mm arises from the molecular roational transitions, mainly the J:1-0,\,2-1,\,3-2 lines. One of the best CO survey to date is provided by 6 . The survey probed the J:1-0 CO line along the Galactic mid-plane with resolution ranging from 7.5’ to 30’. However, the 6 survey was not planned to observe intermediate Galactic latitude regions as it requires longer integration times. On the other hand, Planck released CO emission maps with similar level of accuracy but observations far from the Galactic plane ( |b|>30\deg ) have been shown to be noise dominated ( 22 ) . CO molecular lines are expected to be polarized at lower fractional levels <1% ( 8 ) . 23 , showed how neglecting the polarized emission, mainly from the brightest rotational line of J:1-0 may potentially result into contaminating the polarization measurements of future CMB experiments (e.g. Simons Observatory (SO , 1 ; 12 ) that will further shed light into physics of primordial universe. To date, CO polarized emission has been observed in the brightest regions of molecular cloud complexes and around the Galactic plane, however, SO observations will be instrumental to set more stringent constraints over regions far from the Galactic plane. Because of the intrinsically dim fractional level of CO polarization and the long integration time required to achieve a significant detection, it is difficult to carry out CO polarization surveys on wide areas in the sky ( 6 ) . In the absence of data-based templates, the approach to modeling CO has been to assume a small degree of polarization applied to total intensity maps. Moreover, to cope with the lack of data at high-Galactic latitudes, 23 proposed a model of Galactic molecular clouds to simulate the emission of CO lines in molecular clouds and extrapolate it at high Galactic latitudes. Their model is driven by observational constraints related to the 3D spatial distribution and to the intrinsic properties of the clouds, and it has been already employed in several contexts of Galactic science in the microwaves ( 13 ; 4 ) . However, this approach has the drawback that it yields simulations which, although realistic in themselves, are completely uncorrelated with the other tracers commonly employed to monitor molecular clouds (e.g. HI and HII regions, free–free emission, thermal dust, etc.). Therefore, this work aims at overcoming the limitations presented in 23 by leveraging the latest developments of Generative algorithms in natural image processing. In Section 2 , we present the data-sets and methodology to select the regions of interest, in Section 3 we describe the network architecture employed. Results are discussed in Section 4 and in Section 5 we derive the conclusions and future outlooks. Figure 1: (top panel) Fullsky maps Planck thermal dust at 857 GHz, N(HI) column density map from HI4PI survey and (bottom panel) Planck CO Type 2 J:1-0 and J:2-1 maps are shown respectively from the left to the right column. 2 Data The selection of the training data set is primarily determined by exploiting the observed correlations between molecular emissions, the cold neutral atomic medium and the thermal emission of dust. Although recent literature ( 14 ) indicates that morphological correlations can effectively trace one emission as a function of the other, the relationships between them are highly non-linear and, to date, hard to express with an analytical relation. The Planck satellite has released thermal dust intensity maps with an unprecedented sensitivity in the range of frequencies from 217-857 GHz. We employ the Planck Generalized Needlet Independent Linear Component (GNILC) 2015 component separated map 1 1 1 COM_CompMap_Dust-GNILC-F857_2048_R2.00.fits at 857 GHz ( 21 ) from the Data-release 3. The map was obtained at the angular resolution of 5^{\prime} with the GNILC method aimed at optimally separating the thermal Galactic emission from the extra-galactic contribution due to the CMB and the Cosmic Infrared Background (CIB). We employ the GNILC map at 857 GHz as it is the one with the highest SNR among the Planck maps. As molecular emission arises in the densest filaments and clouds emitting HI atomic emission, we employed the \rm N({HI}) map from HI4PI survey ( 3 ) obtained by integrating data from the coverage of the Effelsberg-Bonn HI Survey and the Galactic All-Sky Survey (GASS) . The angular resolution ( 16.2^{\prime} ) and the improved sensitivity (43 mK) makes it an unprecedented dataset with complete spatial sampling of Galactic HI. Data have been projected onto a HEALPix ( 10 ) map and made publicly available 2 2 2 https://doi.org/10.26093/cds/vizier.35850041 . Galactic CO line emission has been detected even in the broad photometric Planck bands ( 22 ) . They released the Type-2 templates 3 3 3 HFI_CompMap_CO-Type2_2048_R2.00.fits based on a multi-frequency analysis in order to separate the first two rotational CO lines, J:1-0 and J:2-1 , from the diffuse Galactic emission and the CMB. Although this approach is more prone to modeling errors in low SNR regions, we use these maps since they are employed in high SNR areas. The Type-2 CO templates have a resolution of 15^{\prime} , see Fig. 1 . 3 Cycle-GAN model Figure 2: Overall architecture and workflow of the Cycle-GAN implemented in this work. 9 is a widely used machine learning framework that has proven particularly effective in a variety of applications, most notably in image reconstruction tasks. In its standard formulation, a GAN can be understood as an adversarial interplay between two neural networks: a generator , G , and a discriminator , D . The role of G is to synthesize artificial samples that closely mimic the distribution of the training data, while D is trained to assess each input and decide whether it originates from G or is an authentic image drawn from the training set. The objective function of the GAN is constructed so that G is penalized when it produces images that appear unrealistic or that statistically diverge from the true data distribution, whereas D is penalized when it incorrectly classifies a generated image as real or a real image as fake. Through this adversarial training process, the two networks are iteratively optimized until the generator becomes capable of producing high-quality images that the discriminator can no longer reliably distinguish from genuine samples. Lately, GANs have been applied in a wide range of scenarios and for various purposes, consistently demonstrating their ability to generate realistic astrophysical datasets (see 24 ; 16 ; 27 ; 26 ). Motivated by these successful results, we choose to base our approach in this work on such methodologies. 28 developed a method that can learn to translate images from one style or a “collection” to another without needing paired examples. This is a significant breakthrough in unpaired image-to-image translation, a field that aims to convert an image from one representation to another. Traditional image translation methods rely on supervised learning, where a system is trained on thousands of matching input-output pairs (e.g. a photo and its corresponding grayscale version). On the other hand, unsupervised methods like Generative Adversarial Neural networks (GANs) method does not guarantee a meaningful relationship between the input and the translated output and often leads to mode collapse , where the system would produce a single, generic output image for all inputs. To overcome this, 28 added a key constraint: the cycle consistency , by training two models simultaneously, each one accounting for a generator and an adversarial discriminator. Namely: 1. a generator G_{AB}:A\rightarrow B translating images from domain A to domain B and the discriminator D_{B} judging whether a translated image from G_{AB} looks like a real image from domain B ; 2. a generator G_{BA}:B\rightarrow A translates images from domain B to domain A , coupled with the discriminator D_{A} , ensuring the images generated by G_{BA} look like the ones in domain A . Details on the generator and discriminator architectures can be found in A . In our implementation, we choose as adversarial loss the Mean-Squared Error (MSE) loss instead of the standard binary Cross-Entropy , this is mainly due to the fact that the former has shown to be more stable during training ( 15 ; 28 ) . So that for D_{B} and for G the loss functions are defined as follows: \displaystyle\mathcal{L}_{GAN}(D_{B})= \displaystyle\,\frac{1}{2}\mathbb{E}_{b\sim p_{data}(B)}\left[\parallel D_{B}(b)-1\parallel^{2}\right] (1) \displaystyle+ \displaystyle\frac{1}{2}\mathbb{E}_{a\sim p_{data}(A)}\left[\parallel D_{B}(G_{AB}(a)\parallel^{2}\right], (2) \displaystyle\mathcal{L}_{GAN}(G_{AB})= \displaystyle\,\mathbb{E}_{a\sim p_{data}(A)}\left[\parallel D_{B}(G_{AB}(a))-1\parallel^{2}\right] (3) with p_{data} being the data distribution of the training samples. G_{AB} aims to minimize this loss against the adversarial discriminator D_{B} that instead tries to maximize it. From eq.( 3 ), we can observe that the adversarial training process reaches convergence when the discriminator can no longer differentiate between real images and those produced by the generator. In practical terms, this situation corresponds to the point at which the adversarial loss stabilizes around a value of approximately \sim 0.25 , indicating that the discriminator’s predictions have effectively become indistinguishable for real and generated samples. The adversarial loss for G_{BA} and D_{A} are similarly defined. Moreover, the Cycle-GAN objective function is composed also by other two terms : 1. a cycle-consistency loss \mathcal{L}_{\rm cyc} : \displaystyle\mathcal{L}_{\rm cyc}(G_{AB},G_{BA})= \displaystyle\,\mathbb{E}_{a\sim p_{data}(A)}\left[\,\parallel G_{BA}(G_{AB}(a))-a\parallel_{1}\right] \displaystyle+ \displaystyle\,\mathbb{E}_{b\sim p_{data}(B)}\left[\,\parallel G_{AB}(G_{BA}(b))-b\parallel_{1}\right], ensuring the forward cycle consistency , i.e. for each image a when we apply consecutively G_{AB} and G_{BA} we retrieve back the original image, i.e. a\rightarrow G_{AB}(a)\rightarrow G_{BA}(G_{AB}(a))\approx a. Similarly, for each image b from domain B the backward cycle consistency grants the reverse process. We adopt L_{1} norm for \mathcal{L}_{\rm cyc} . 2. an identity loss, \mathcal{L}_{\rm id} : \displaystyle\mathcal{L}_{\rm id}(G_{AB},G_{BA})= \displaystyle\,\mathbb{E}_{a\sim p_{data}(A)}\left[\,\parallel G_{BA}(a)-a\parallel_{1}\right] \displaystyle+ \displaystyle\,\mathbb{E}_{b\sim p_{data}(B)}\left[\,\parallel G_{AB}(b)-b\parallel_{1}\right], ensuring that if a generator is given an image already belonging to the target domain, it should leave it unchanged. The relative importance of the two objectives is controlled by two hyper-parameters, as it follows: \displaystyle\mathcal{L}_{\rm total} \displaystyle=\mathcal{L}_{GAN}(G_{AB})+\mathcal{L}_{GAN}(G_{BA}) \displaystyle+\lambda_{\rm cyc}\mathcal{L}_{\rm cyc}(G_{AB},G_{BA}) \displaystyle+\lambda_{\rm id}\mathcal{L}_{\rm id}(G_{BA},G_{BA}), with \lambda_{\rm cyc}=10 and \lambda_{\rm id}=5 (values chosen following the recommendations from 28 ). Further details on the architecture of the generator and discriminatror can be found A . 3.1 Training To construct the dataset and facilitate the training process, we convolve the four input maps to the coarsest angular resolution of 16^{\prime} and make sure that all the maps are pixellated following the HEALPix scheme, with the same pixel resolution of \sim 1.7^{\prime} , corresponding to an nside =2048 HEALPix grid. One of the advantages of GANs is to employ unpaired data sets, i.e. training with an intentionally mismatched data from A and B domains, in order to let the network learns the domain translation rules, not just how to copy one specific image to another. This is particularly remarkable when dealing with realistic datasets, where observations could be highly affected by noise or could involve different kinds of emissions. As can be seen in Fig. 1 , both the Planck 857 GHz dust and the HI4PI maps are signal-dominated, with a signal-to-noise ratio (SNR) > 3 across the full-sky coverage ( 20 ; 3 ) . This high SNR ensures that the features we use as inputs are reliably measured and noise, albeit present, is subdominant with respect to the signal. To construct the input data, we therefore segmented the Planck and HI4PI all-sky maps into tiles of size 3\times 3\deg^{2} , each tile consisting of 128\times 128 pixels, accounting for a total of 8,452 patches. On the opposite, the Planck CO Type-2 maps (the output targets) are more contaminated by noise especially far from the Galactic mid-plane. We, thus, consider 3\times 3\deg^{2} tiles, made of 128\times 128 pixels, chosen from SNR> 8 regions as illustrated in Fig. 3 . The dataset is further augmented through flipping, aiming to further enhance the size and diversity of the training dataset, leading to a total of 5,205 (whose 336 left for testing) and 22,818 respectively for the output and the input maps. We left 336 paired tiles for testing to allow a more direct comparison of the performances. We remark here that the SNR >8 selection of regions may correspond to systematically select optically thick regions, notably near the plane. We, thus,make sure to construct the whole training data set including also optically thin regions at intermediate galactic latitude or in the outer galaxy, allowing the network to inspect those emission regimes. In fig. 3 , we show in different colors the regions selected for training, validation and test. We notice that each data set employs tiles selected from the Galactic plane and from the outer Galaxy or intermediate Galactic latitudes. Additionally, we mitigated any overlap between tiles to preserve morphological diversity within each set. This leads to a total of 1,733 tiles (whose locations are shown in Fig. 3 ). Figure 3: High SNR regions of CO emission from the Planck Type-2 maps. Different colors refers to patches employed for (orange) training, (red) validation, (brown) test sets. To compress the dynamic range, we firstly log- transform each HEALPix map, M , i.e. : m=\log_{10}(M+\epsilon), with \epsilon=10^{-6} , appropriately chosen to avoid NaN errors with zero values in M . We also make sure to shift to 0 any negative values due to instrumental noise. We, then, minmax normalized the map so that it globally ranges in the interval [-1,1] : \tilde{m}=2\frac{(m-\min(m))}{\max(m)-\min(m)}-1. We, finally, proceed to build the training set by projecting into squared patches from the normalized map \tilde{m} . We consider batches of 32 tiles and perform the training onto an NVIDIA A100-SXM4-40GB GPU node of the supercomputing facility PERLMUTTER at NERSC. The optimizer adopted is the Adaptive Moment Estimation (ADAM) optimizer with a learning rate of 0.0002 and \beta parameters set to (0.5,0.999) . Figure 4: (top) Generator cycle (solid green) and identity (dotted blue) losses as a function of epochs. (bottom) Discriminator loss as a function of epochs. The vertical dashed thick line at epoch 460 marks the point where the training switched from paired to unpaired. The training employes 940 epochs and take \sim 47 hours (each one taking on average 3 min, with a 0.1% spent for data loading, and 14% for batch processing time). We decide to perform the training with a paired data set to allow the network learning astrophysical correlations present in both input and output features. As shown in Fig. 4 , , we notice that after few hundred epochs, the generator loss starts to increase and reaches a plateau of 0.03 after 460 epochs. This is expected since the paired data set is by construction smaller than the unpaired one (due to the limited SNR>8 CO regions). We, thus, decide to switch into unpaired training by further learning the features in both optically thin and optically thick regimes. During unpaired training, we increased also \lambda_{\rm cyc}=20 , to give more weight to the cycle loss. This results in a sensible decrease of the cycle loss until it reaches the desired threshold of convergence at 0.02 . The code has been made publicly available 4 4 4 https://github.com/giuspugl/COnet.git together with the training weights 5 5 5 Cycle-GAN weights for the sake of reproducibility. 4 Results Figure 5: 3\times 3\deg^{2} tiles employed in the test set for training the Cycle-GAN selected from regions shown in Fig. 3 . Planck thermal dust at 857 GHz, N(\rm{HI}) column density map from HI4PI survey, Planck CO Type 2 J:1-0 and J:2-1 maps are shown respectively from the left to the right column, together with the predictions obtained with the Cycle-GAN . In Fig. 5 , we show a visual comparison of the patches generated with the Cycle-GAN against the ground truth in the test set. We notice that the network is able to reproduce similar filamentary and point-source structures as solely inspecting the input features represented by the dust and HI maps from the test set. We can clearly state that the network consistently predicts not only realistic morphological shapes but also coherent range of amplitudes. To better guide the reader, we set the colorbars of the mock images to be the same as the real ones. Moreover, we notice the morphologies are mainly guided by the the thermal dust input features. This outcome is, to some extent, expected, as astrophysical observations indicate that molecular emission closely correlates with the emission from cold dust. ( 25 ) . We then proceed with a more quantitative estimate by means of the 2-point statistics, to assess predicted data quality in terms of angular correlations in the synthesized images by estimating the angular power spectrum of each tile of the test set using NaMaster software ( 2 ) . For most of Galactic emission, we expect the power spectrum to be scale-invariant, following a power-law scaling in terms of the multipole \ell as shown in Fig. 6 . As our patches involves a small portion of the sky, where curvature effects are negligible, we can run NaMaster with Flat-Field estimators as recommended in the code documentation 6 6 6 https://namaster.readthedocs.io/en/latest/api/pymaster.utils.html . We further consider a window function to apodize each tile with the C2 profile, defined as follows: w(x)=\begin{cases}\frac{1}{2}\left(1-\cos(\pi x)\right),&\text{if }x<1\\ 1&\mathrm{otherwise}\end{cases} with x=\sqrt{(1-\cos\theta)/(1-\cos\theta_{*})} , \theta being the distance between two pixel and \theta_{*}=0.5\deg . The power spectra are subsequently estimated with equally spaced bins of multipoles characterized by \Delta\ell=200 . To facilitate the comparison of spectra derived from maps with differing units of measure, we estimate the spectra directly on the log-normalized data set, as described in Subsection 3.1 . Figure 6: Power spectra estimated from the test set. Median power spectra are shown as (solid orange) and (solid green) respectively for the ones estimated from real and synthesized (mock) CO J:1-0 (left panel) and J:2-1 (right panel) emissions. The green and orange shaded areas corresponds to the 2\sigma interval of power spectra estimated respectively from the whole mock and real test set. Power spectra of input dust and HI maps are respectively shown for comparison in (dot-dashed blue) and (dot-dashed red). The spectra are uniformly binned with \Delta\ell=200 . Fig. 6 shows the power spectra estimated from the test set. We notice that the median (solid green ) from Cycle-GAN CO predictions almost follows the same scaling as the ground-truth one (solid orange). We remark here that the estimated power spectra accurately reproduce the scaling across various angular scales, ranging from 0.5\deg to 3^{\prime} . Moreover, we clearly are able to distinguish that the network is reproducing more the scaling observed in the CO power spectrum than the dust and HI ones. We also show the 2\sigma excursion range of the power spectra estimated from the different tiles of the predictions and ground-truth maps. As both the excursion range overlaps in most of the angular multipoles, we can quantitatively assess that the angular correlations of the mock and real data are consistent withing the 2\sigma interval. Figure 7: From left to right, {V_{0},V_{1},V_{2}} MFs estimated from the test set for the J:1-0 and J:2-1 CO median emission from the ground-truth (solid orange) compared with the mock one syntethized by the Cycle-GAN (solid green). As in Fig. 6 , we also show the 2\sigma interval of the excursion range as shaded orange and green respectively for the ground-truth and mock sets. For the sake of comparison we also show the median MFs of the input dust (dot-dashed blue) and the log of HI column density (dot-dashed red). Figure 8: 3\times 3\deg^{2} tiles employed for application of the Cycle-GAN at high-Galactic latitudes. From left to right columns we show maps (in log-normalized units) : Planck thermal dust at 857 GHz, N(\rm{HI}) column density map from HI4PI survey, CO J:1-0 and J:2-1 from mock, Planck Type 2 ,and pysm3 model maps. To facilitate the comparison between the CO emission maps, we set their colorbars to the same range. We further assess through the Minkowski Functionals ( 19 , MFs,) the degree of non-Gaussianity in the mock CO emission compared with the ground-truth one. MF have been widely employed in the literature to estimate higher-order statistical properties of maps ( 24 ; 17 ; 18 ; 7 ; 5 ; 27 ; 4 , e.g.,) . For a given n -dim set whose range is defined with a threshold \rho , there are n+1 MFs that could provide high-order statistics informations ( 11 ) . In this work we employ n=2 maps, thus we utilize 3 MFs: {V}_{0} , {V}_{1} , and {V}_{2} , representing respectively the area, perimeter, and connectivity as a function of the threshold \rho . We estimate MFs with the package pynkowski 7 7 7 https://github.com/javicarron/pynkowski ( 7 ) . Fig. 7 shows the MFs estimated from the patches in the test set, once they are standardized (zero mean and variance equal to 1). We employ the same color scheme as in Fig 6 : the mock (solid green) and ground-truth (solid orange) emissions, accounting for the 2\sigma interval of the excursion range respectively shown in shaded green and orange. We remark that the MFs are consistent within the 2\sigma interval, indicating that also the high-order statistics has been properly learnt and reproduced by the Cycle-GAN . By looking at the MFs estimated dust and HI patches, we clearly see a remarkable difference when compared with the ones estimated from mock data. This further points to the fact that the high-order statistics of generated mock data reproduces the one observed in real ones. 4.1 Application at high-Galactic latitudes In this Sub-section, we assess the performances of our trained Cycle-GAN to generalize in regions at high-Galactic latitudes, where current CO data are currently lacking or are strongly affected by noise. We therefore divide the full-sky dust and HI observations into 8,452 3\times 3\,\mathrm{deg^{2}} tiles performing the same preprocessing steps followed in Sub-section 3.1 and run the Cycle-GAN to predict on this data set. In Fig. 8 , we show a subset of these predictions, selected from all the patches at the same Galactic longitude \ell=240^{\circ} and at several Galactic latitudes. This choice is specifically driven by the fact that the training set does not encode data along this longitude, as can be noticed from Fig. 3 . We compare the predictions against the Planck CO Type-2 and we clearly state that for all the patches shown in Fig. 8 the latter are strongly contaminated by instrumental noise in both the two CO lines, As stated above, we aim to compare our predictions against the Monte-Carlo Molecular 3D ( MCMole3D ) model from 23 . Their model populated high Galactic latitude CO clouds by fitting several parameters from low latitudes CO emission data. A notable distinction from this approach is that they use Planck CO Type-1 maps as templates for the emission of both J:1-0 and J:2-1 lines. Recently, the MCMole3D maps have been incorporated into the Python Sky Model 3 ( pysm3 , 4 ) 8 8 8 See the model labeled as co3 in the pysm3 documentation . . We, therefore, run the pysm3 to simulate the CO J:1-0 and J:2-1 emission from the co3 model, log-normalized and projected into square tiles. When comparing our mock maps with the pysm3 ones, we clearly notice several differences among the two data sets. On one hand, being the pysm3 model based on the Planck CO Type-1 maps, its emission is largely contaminated by the instrumental noise. This is expected, as the CO emission is expected to be weaker at high-Galactic latitudes, making instrumental sensitivity a critical factor. Indeed, the Cycle-GAN predicts pixel values of -1 in the areas, that are likely to be associated with intrinsically faint emission and more prone to be contaminated by instrumental noise. On the other hand, the morphology of the emission in our predicted maps more closely reproduces the diffuse, filamentary structures observed in dust and HI, whereas in the pysm3 maps, residual noise readily distorts the morphology, favoring compact, point-like emission features at the expense of the extended structures. 5 Conclusions In this work, we introduce a novel methodology utilizing a Cycle-GAN implementation to predict realistic non-Gaussian sub-degree CO emissions. Our algorithm is employed to learn molecular CO emissions by examining inputs from thermal dust emissions and atomic emissions while targeting CO emissions in the two most prominent rotational transitions, J:1-0 and J:2-1 . The network is trained to generate CO emissions specifically in regions where observations have been achieved with SNR>8. The data set encodes square sky regions with a physical size of 3\times 3\deg^{2} and a resolution of 128\times 128 pixels. The quality of the results, quantified through the power spectra and MFs, demonstrate that our algorithm successfully meets the aim of this work, illustrating that the angular correlations of the generated features properly scale with angular size and have the same high-order statistics as the CO targets employed. This has been evaluated by assessing consistency within 2\sigma with a comparison of power spectra and MFs of the mock and real data sets. To further corroborate the performance of our method, we show that that our methodology is able to generalize at high Galactic latitudes which have not been taken into account in the training set. This is particular of interest as molecular complexes far from the Galactic plane are characterized by low emissivity and low opacity, making surveys difficult to carry out in those regions. We further compare the predictions with a model from the pysm3 package commonly adopted by the community. From such a comparison, we expect differences not only in the amplitudes but also in the morphologies, because the pysm3 model is more affected by the Planck instrumental noise, whereas the model presented in this work completely relies on high SNR of thermal dust and HI maps. This work paves the road for a new Galactic CO emission model at high Galactic latitudes that will be presented in a future work. More remarkably, it demonstrates the advantages of employing such algorithms to cope the current limitations arising from the available observational data sets. Appendix A Cycle-GAN Architecture The architecture shown in Fig. 2 consists of two generator models and two discriminators. The cycle consistency loss is the core component that ensures successful transformation from one domain to another and back again, thus maintaining the original content. We remark here, that we employ the exact same architecture for both discriminators D_{X} and D_{Y} and for the generators, G and F . The generator (Fig. 9 ) employs a ResNet-based Encoder-Decoder structure. The Encoder stage is made of 3 blocks of 2D convolutional layers, with a ReLU activation function, aimed at downsampling the image to extract "features" while discarding spatial redundancy. At the end of the Encoder stage, there are 6 Residual blocks with skip-connections , so that the input in each block is added to the output and with a ReLU activation function. Finally, the Decoder reconstructs processed features back into an image, via 3 consecutive transposed convolutional layers (or Deconvolutions). The output layer maps the features back to the output channels, followed by a \tanh activation function to force the values to range into the [-1,1] range. Moreover, we employ Instance Normalization that normalizes each image independently enhancing the absolute contrast of each specific map. We choose the PatchGAN architecture (Fig. 10 ) for the discriminators, specifically used in Cycle-GAN (and Pix2Pix architectures ( 15 ) ). A PatchGAN outputs a grid (in our case 14\times 14 ) of probabilities representing how much is likely to look real a specific patch in the input image. It is a fully convolutional network, and forces the Generator to produce high-frequency details, which are frequently lost in the context of image generation. We employ LeakyReLU activation function adjusted to the slope of 0.2 . Figure 9: Cycle-GAN Generator architecture. (inset) a zoom-in showing the layers composing each Residual Block. Figure 10: Cycle-GAN Discriminator architecture. Figure 11: Performance of the G_{BA} generator, in this case the input features are the CO 1-0 and 2-1 maps (left column), the generated and the output targets of thermal dust and N(HI) emissions, respectively in middle and right column.