Compact Muon Solenoid
LHC, CERN
Visit us: CMS Public Website, CMS Physics ;
Contact us: CMS Publications Committee
| CMS-PAS-TAU-24-001 | ||
| Identification of tau leptons using a convolutional neural network with domain adaptation in the CMS experiment | ||
| CMS Collaboration | ||
| 3 May 2025 | ||
| Abstract: The DeepTau identification algorithm, based on deep neural network techniques, has been developed to reduce the fraction of jets, muons, and electrons misidentified as hadronically decaying tau leptons ($ \tau_\mathrm{h} $) in the CMS experiment. The latest version of this algorithm includes domain adaptation by backpropagation, a technique that reduces data-to-simulation discrepancies in the region with the highest purity of genuine $ \tau_\mathrm{h} $ candidates. Additionally, a refined training workflow improves classification performance, with a reduction of 30-50% in the probability for jets to be misidentified as a $ \tau_\mathrm{h} $ for a given reconstruction and identification efficiency. This note presents the main novelties introduced to the DeepTau algorithm and evaluates its performance in LHC proton-proton collision data at $ \sqrt{s}= $ 13 and 13.6 TeV collected in 2018 and 2022, respectively, with integrated luminosities of 60 and 35 fb$ ^{-1} $. The techniques to determine data-to-simulation scale factors are presented with a subset of results among the ones deployed centrally for CMS physics analyses. | ||
| Links: CDS record (PDF) ; CADI line (restricted) ; | ||
| Figures | |
png pdf |
Figure 1:
Schematic illustration of the signatures of the $ \mathrm{h}^{\pm} $, $ \mathrm{h}^{\pm}\pi^{0} $, $ \mathrm{h}^{\pm}\mathrm{h}^{\mp}\mathrm{h}^{\pm} $, and $ \mathrm{h}^{\pm}\mathrm{h}^{\mp}\mathrm{h}^{\pm}\pi^{0} $ decay modes of the tau lepton in the CMS detector. Charged hadrons are reconstructed by the PF algorithm by matching tracks with energy deposits in the ECAL and HCAL, while the HPS algorithm aims to reconstruct each $ \pi^{0}\to\gamma\gamma $ decay as a single ``strips'' of energy clusters in ECAL. |
png pdf |
Figure 2:
Inner and outer grid layout in $ \eta $-$ \phi $ space [21]. The inner grid encapsulates the signal cone of maximal radius 0.1, which contains the $ \mathrm{h}^{\pm} $ and $ \pi^{0} $ candidates, and consists of 11 $ \times $ 11 cells with a size of 0.02 $ \times $ 0.02 each. The outer grid contains the isolation cone of radius 0.5, and consists of 21 $ \times $ 21 cells with a size of 0.05 $ \times $ 0.05 each. |
png pdf |
Figure 3:
The DEEPTAU architecture with the domain adaptation configuration [66]. A set of final domain adaptation layers was introduced for data-simulation discrimination, consisting of several dense layers followed by a softmax layer that yields an output $ y_\text{adv} $ between zero and one. The backpropagation is modified to include the ``adversarial loss'', as described in the text. |
png pdf |
Figure 4:
Distribution of the DEEPTAU discriminator against jets before (left) and after (right) domain adaptation, for the 2018 dataset used for domain adaptation training. There is significant improvement in data-simulation agreement in the signal region, with the discrepancies in the final bin reduced to 0.9%. |
png pdf |
Figure 4-a:
Distribution of the DEEPTAU discriminator against jets before (left) and after (right) domain adaptation, for the 2018 dataset used for domain adaptation training. There is significant improvement in data-simulation agreement in the signal region, with the discrepancies in the final bin reduced to 0.9%. |
png pdf |
Figure 4-b:
Distribution of the DEEPTAU discriminator against jets before (left) and after (right) domain adaptation, for the 2018 dataset used for domain adaptation training. There is significant improvement in data-simulation agreement in the signal region, with the discrepancies in the final bin reduced to 0.9%. |
png pdf |
Figure 5:
Distribution of the DEEPTAU discriminator against jets before (left) and after (right) domain adaptation, for the early 2022 dataset. While data-to-simulation differences remain, there is an appreciable improvement with the inclusion of domain adaptation. |
png pdf |
Figure 5-a:
Distribution of the DEEPTAU discriminator against jets before (left) and after (right) domain adaptation, for the early 2022 dataset. While data-to-simulation differences remain, there is an appreciable improvement with the inclusion of domain adaptation. |
png pdf |
Figure 5-b:
Distribution of the DEEPTAU discriminator against jets before (left) and after (right) domain adaptation, for the early 2022 dataset. While data-to-simulation differences remain, there is an appreciable improvement with the inclusion of domain adaptation. |
png pdf |
Figure 6:
Jet misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated on 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H}\to \tau \tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Jet misidentification probability is estimated from $ {\mathrm{t}\overline{\mathrm{t}}} $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that do not match prompt electrons, muons or products of $ \tau_\mathrm{h} $ decays at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 6-a:
Jet misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated on 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H}\to \tau \tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Jet misidentification probability is estimated from $ {\mathrm{t}\overline{\mathrm{t}}} $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that do not match prompt electrons, muons or products of $ \tau_\mathrm{h} $ decays at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 6-b:
Jet misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated on 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H}\to \tau \tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Jet misidentification probability is estimated from $ {\mathrm{t}\overline{\mathrm{t}}} $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that do not match prompt electrons, muons or products of $ \tau_\mathrm{h} $ decays at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 7:
Electron misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated on 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H} \to \tau\tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Electron misidentification probability is estimated from Drell--Yan simulation using reconstructed $ \tau_\mathrm{h} $ candidates that match electrons at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 7-a:
Electron misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated on 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H} \to \tau\tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Electron misidentification probability is estimated from Drell--Yan simulation using reconstructed $ \tau_\mathrm{h} $ candidates that match electrons at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 7-b:
Electron misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated on 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H} \to \tau\tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Electron misidentification probability is estimated from Drell--Yan simulation using reconstructed $ \tau_\mathrm{h} $ candidates that match electrons at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 8:
Muon misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H} \to \tau\tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Muon misidentification probability is estimated from Drell--Yan simulation using reconstructed $ \tau_\mathrm{h} $ candidates that match muons at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 8-a:
Muon misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H} \to \tau\tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Muon misidentification probability is estimated from Drell--Yan simulation using reconstructed $ \tau_\mathrm{h} $ candidates that match muons at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 8-b:
Muon misidentification probability versus $ \tau_\mathrm{h} $ identification efficiency for low-$ p_{\mathrm{T}} $ (left) and high-$ p_{\mathrm{T}} $ (right) $ \tau_\mathrm{h} $ candidates, evaluated 2018 simulated datasets. The $ \tau_\mathrm{h} $ identification efficiency is estimated from $ \mathrm{H} \to \tau\tau $ simulations using reconstructed $ \tau_\mathrm{h} $ candidates that match generator level $ \tau_\mathrm{h} $ objects. Muon misidentification probability is estimated from Drell--Yan simulation using reconstructed $ \tau_\mathrm{h} $ candidates that match muons at the generator level. The defined working points of the discriminator are indicated as filled circles. |
png pdf |
Figure 9:
Distribution of the visible mass in the $ \mu\tau_\mathrm{h} $ channel when using DEEPTAU v2.1 (left) and v2.5 (right) for discrimination in the 2018 dataset. |
png pdf |
Figure 9-a:
Distribution of the visible mass in the $ \mu\tau_\mathrm{h} $ channel when using DEEPTAU v2.1 (left) and v2.5 (right) for discrimination in the 2018 dataset. |
png pdf |
Figure 9-b:
Distribution of the visible mass in the $ \mu\tau_\mathrm{h} $ channel when using DEEPTAU v2.1 (left) and v2.5 (right) for discrimination in the 2018 dataset. |
png pdf |
Figure 10:
The data-to-simulation scale factors of the $ \tau_\mathrm{h} $ identification efficiency as a function of $ p_{\mathrm{T}} $ in the 2018 (left) and 2022 (right) data-taking periods, including all $ \tau_\mathrm{h} $ decay modes, and requiring the $ D_\text{jet} $ Medium working point and an $ m_{\mathrm{T}}(\mu,p_{\mathrm{T}}^\text{miss}) < $ 65 GeV cut. The vertical bars correspond to the combined statistical and systematic uncertainties in the individual scale factors. |
png pdf |
Figure 10-a:
The data-to-simulation scale factors of the $ \tau_\mathrm{h} $ identification efficiency as a function of $ p_{\mathrm{T}} $ in the 2018 (left) and 2022 (right) data-taking periods, including all $ \tau_\mathrm{h} $ decay modes, and requiring the $ D_\text{jet} $ Medium working point and an $ m_{\mathrm{T}}(\mu,p_{\mathrm{T}}^\text{miss}) < $ 65 GeV cut. The vertical bars correspond to the combined statistical and systematic uncertainties in the individual scale factors. |
png pdf |
Figure 10-b:
The data-to-simulation scale factors of the $ \tau_\mathrm{h} $ identification efficiency as a function of $ p_{\mathrm{T}} $ in the 2018 (left) and 2022 (right) data-taking periods, including all $ \tau_\mathrm{h} $ decay modes, and requiring the $ D_\text{jet} $ Medium working point and an $ m_{\mathrm{T}}(\mu,p_{\mathrm{T}}^\text{miss}) < $ 65 GeV cut. The vertical bars correspond to the combined statistical and systematic uncertainties in the individual scale factors. |
png pdf |
Figure 11:
The $ m_\text{vis} $ distribution in the $ \mathrm{Z}\to\tau_\mu\tau_\mathrm{h} $ channel for year 2022 before (left) and after (right) the full calibration. The DEEPTAU working points used are: Medium for $ D_\text{jet} $, VVLoose for $ D_\mathrm{e} $ and, Tight for $ D_\mu $. The application of correction factors improves the agreement between data and simulation. |
png pdf |
Figure 11-a:
The $ m_\text{vis} $ distribution in the $ \mathrm{Z}\to\tau_\mu\tau_\mathrm{h} $ channel for year 2022 before (left) and after (right) the full calibration. The DEEPTAU working points used are: Medium for $ D_\text{jet} $, VVLoose for $ D_\mathrm{e} $ and, Tight for $ D_\mu $. The application of correction factors improves the agreement between data and simulation. |
png pdf |
Figure 11-b:
The $ m_\text{vis} $ distribution in the $ \mathrm{Z}\to\tau_\mu\tau_\mathrm{h} $ channel for year 2022 before (left) and after (right) the full calibration. The DEEPTAU working points used are: Medium for $ D_\text{jet} $, VVLoose for $ D_\mathrm{e} $ and, Tight for $ D_\mu $. The application of correction factors improves the agreement between data and simulation. |
png pdf |
Figure 12:
The $ m_\text{vis} $ distribution in the $ \mathrm{Z}\to\tau_\mathrm{e}\tau_\mathrm{h} $ channel for year 2022 before (left) and after (right) the full calibration. The DEEPTAU working points used are: Medium for $ D_\text{jet} $, Tight for $ D_\mathrm{e} $ and, Tight for $ D_\mu $. Specific 2022 detector conditions that affected electron reconstruction are not perfectly modelled in the simulation. As a result, the amount of electrons misidentified as $ \tau_\mathrm{h} $ is enhanced in data with respect to simulated events. The application of correction factors improves the agreement between data and simulation. |
png pdf |
Figure 12-a:
The $ m_\text{vis} $ distribution in the $ \mathrm{Z}\to\tau_\mathrm{e}\tau_\mathrm{h} $ channel for year 2022 before (left) and after (right) the full calibration. The DEEPTAU working points used are: Medium for $ D_\text{jet} $, Tight for $ D_\mathrm{e} $ and, Tight for $ D_\mu $. Specific 2022 detector conditions that affected electron reconstruction are not perfectly modelled in the simulation. As a result, the amount of electrons misidentified as $ \tau_\mathrm{h} $ is enhanced in data with respect to simulated events. The application of correction factors improves the agreement between data and simulation. |
png pdf |
Figure 12-b:
The $ m_\text{vis} $ distribution in the $ \mathrm{Z}\to\tau_\mathrm{e}\tau_\mathrm{h} $ channel for year 2022 before (left) and after (right) the full calibration. The DEEPTAU working points used are: Medium for $ D_\text{jet} $, Tight for $ D_\mathrm{e} $ and, Tight for $ D_\mu $. Specific 2022 detector conditions that affected electron reconstruction are not perfectly modelled in the simulation. As a result, the amount of electrons misidentified as $ \tau_\mathrm{h} $ is enhanced in data with respect to simulated events. The application of correction factors improves the agreement between data and simulation. |
png pdf |
Figure 13:
Summary of $ \tau_\mathrm{h} $ energy scales (right) and $ \tau_\mathrm{h} $ identification efficiency (left) across $ \tau_\mathrm{h} $ decay modes and $ p_{\mathrm{T}} $ regions for 2018 with $ m_{\mathrm{T}}(\mu,p_{\mathrm{T}}^\text{miss}) < $ 65 GeV and the $ D_\text{jet} $ Medium working point. |
png pdf |
Figure 13-a:
Summary of $ \tau_\mathrm{h} $ energy scales (right) and $ \tau_\mathrm{h} $ identification efficiency (left) across $ \tau_\mathrm{h} $ decay modes and $ p_{\mathrm{T}} $ regions for 2018 with $ m_{\mathrm{T}}(\mu,p_{\mathrm{T}}^\text{miss}) < $ 65 GeV and the $ D_\text{jet} $ Medium working point. |
png pdf |
Figure 13-b:
Summary of $ \tau_\mathrm{h} $ energy scales (right) and $ \tau_\mathrm{h} $ identification efficiency (left) across $ \tau_\mathrm{h} $ decay modes and $ p_{\mathrm{T}} $ regions for 2018 with $ m_{\mathrm{T}}(\mu,p_{\mathrm{T}}^\text{miss}) < $ 65 GeV and the $ D_\text{jet} $ Medium working point. |
png pdf |
Figure 14:
Summary of $ \tau_\mathrm{h} $ energy scales across $ \tau_\mathrm{h} $ decay modes for 2022 with $ m_{\mathrm{T}}(\mu,p_{\mathrm{T}}^\text{miss}) < $ 40 GeV and $ D_\text{jet} $ Medium working point. |
png pdf |
Figure 15:
Muon misidentification rate scale factors binned by $ |\eta| $ for the Medium $ D_\mu $ working point. Measurement for the 2018 dataset is shown on the left and for the 2022 dataset on the right. The dashed lines indicate the boundaries of $ \tau_\mathrm{h} |\eta| $ bins. |
png pdf |
Figure 15-a:
Muon misidentification rate scale factors binned by $ |\eta| $ for the Medium $ D_\mu $ working point. Measurement for the 2018 dataset is shown on the left and for the 2022 dataset on the right. The dashed lines indicate the boundaries of $ \tau_\mathrm{h} |\eta| $ bins. |
png pdf |
Figure 15-b:
Muon misidentification rate scale factors binned by $ |\eta| $ for the Medium $ D_\mu $ working point. Measurement for the 2018 dataset is shown on the left and for the 2022 dataset on the right. The dashed lines indicate the boundaries of $ \tau_\mathrm{h} |\eta| $ bins. |
png pdf |
Figure 16:
Summary plots of results for electron misidentification rate scale factors divided in decay modes and $ \eta $ regions for the VVLoose $ D_\mathrm{e} $ working point. The corrections are shown on the left for 2018 and on the right for 2022 datasets. |
png pdf |
Figure 16-a:
Summary plots of results for electron misidentification rate scale factors divided in decay modes and $ \eta $ regions for the VVLoose $ D_\mathrm{e} $ working point. The corrections are shown on the left for 2018 and on the right for 2022 datasets. |
png pdf |
Figure 16-b:
Summary plots of results for electron misidentification rate scale factors divided in decay modes and $ \eta $ regions for the VVLoose $ D_\mathrm{e} $ working point. The corrections are shown on the left for 2018 and on the right for 2022 datasets. |
png pdf |
Figure 17:
The high-$ p_{\mathrm{T}} \tau_\mathrm{h} $ identification efficiency scale factors as a function of $ \tau_\mathrm{h} p_{\mathrm{T}} $ for $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ VVLoose (left) and tight (right) discriminators. The scale factors are measured for the 2018 (top) and 2022 (bottom) dataset. |
png pdf |
Figure 17-a:
The high-$ p_{\mathrm{T}} \tau_\mathrm{h} $ identification efficiency scale factors as a function of $ \tau_\mathrm{h} p_{\mathrm{T}} $ for $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ VVLoose (left) and tight (right) discriminators. The scale factors are measured for the 2018 (top) and 2022 (bottom) dataset. |
png pdf |
Figure 17-b:
The high-$ p_{\mathrm{T}} \tau_\mathrm{h} $ identification efficiency scale factors as a function of $ \tau_\mathrm{h} p_{\mathrm{T}} $ for $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ VVLoose (left) and tight (right) discriminators. The scale factors are measured for the 2018 (top) and 2022 (bottom) dataset. |
png pdf |
Figure 17-c:
The high-$ p_{\mathrm{T}} \tau_\mathrm{h} $ identification efficiency scale factors as a function of $ \tau_\mathrm{h} p_{\mathrm{T}} $ for $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ VVLoose (left) and tight (right) discriminators. The scale factors are measured for the 2018 (top) and 2022 (bottom) dataset. |
png pdf |
Figure 17-d:
The high-$ p_{\mathrm{T}} \tau_\mathrm{h} $ identification efficiency scale factors as a function of $ \tau_\mathrm{h} p_{\mathrm{T}} $ for $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ VVLoose (left) and tight (right) discriminators. The scale factors are measured for the 2018 (top) and 2022 (bottom) dataset. |
png pdf |
Figure 18:
Prefit (left plots) and postfit (right plots) distribution of $ m_{\mathrm{T}}(\tau_\mathrm{h}, p_{\mathrm{T}}^\text{miss}) $ for $ p_{\mathrm{T}} $ bins of 100 $ < p_{\mathrm{T}} < $ 200 GeV (upper plots) and $ p_{\mathrm{T}} > $ 200 GeV (lower plots) measurement regions of 2022 dataset. Distributions are obtained for a combination of $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ Tight discriminators. |
png pdf |
Figure 18-a:
Prefit (left plots) and postfit (right plots) distribution of $ m_{\mathrm{T}}(\tau_\mathrm{h}, p_{\mathrm{T}}^\text{miss}) $ for $ p_{\mathrm{T}} $ bins of 100 $ < p_{\mathrm{T}} < $ 200 GeV (upper plots) and $ p_{\mathrm{T}} > $ 200 GeV (lower plots) measurement regions of 2022 dataset. Distributions are obtained for a combination of $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ Tight discriminators. |
png pdf |
Figure 18-b:
Prefit (left plots) and postfit (right plots) distribution of $ m_{\mathrm{T}}(\tau_\mathrm{h}, p_{\mathrm{T}}^\text{miss}) $ for $ p_{\mathrm{T}} $ bins of 100 $ < p_{\mathrm{T}} < $ 200 GeV (upper plots) and $ p_{\mathrm{T}} > $ 200 GeV (lower plots) measurement regions of 2022 dataset. Distributions are obtained for a combination of $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ Tight discriminators. |
png pdf |
Figure 18-c:
Prefit (left plots) and postfit (right plots) distribution of $ m_{\mathrm{T}}(\tau_\mathrm{h}, p_{\mathrm{T}}^\text{miss}) $ for $ p_{\mathrm{T}} $ bins of 100 $ < p_{\mathrm{T}} < $ 200 GeV (upper plots) and $ p_{\mathrm{T}} > $ 200 GeV (lower plots) measurement regions of 2022 dataset. Distributions are obtained for a combination of $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ Tight discriminators. |
png pdf |
Figure 18-d:
Prefit (left plots) and postfit (right plots) distribution of $ m_{\mathrm{T}}(\tau_\mathrm{h}, p_{\mathrm{T}}^\text{miss}) $ for $ p_{\mathrm{T}} $ bins of 100 $ < p_{\mathrm{T}} < $ 200 GeV (upper plots) and $ p_{\mathrm{T}} > $ 200 GeV (lower plots) measurement regions of 2022 dataset. Distributions are obtained for a combination of $ D_\text{jet} $ Medium, $ D_\mu $ Tight and $ D_\mathrm{e} $ Tight discriminators. |
| Tables | |
png pdf |
Table 1:
Selection requirements for the domain adaptation dataset. The impact parameters, $ d_z $ and $ d_{xy} $, are defined as the distances between the muon track (or leading charged hadron track) and the PV for the muon (or $ \tau_\mathrm{h} $ candidate respectively). The medium muon identification is defined in Ref. [27]. The previous DEEPTAU discriminator scores described in Ref. [21] against quark and gluon jets, electrons, and muons, are denoted $ D_\text{jet}^\text{v2.1} $, $ D_\mathrm{e}^\text{v2.1} $, and $ D_\mu^\text{v2.1} $. The transverse mass of the muon and the missing transverse energy is denoted as $ m_{\mathrm{T}}(\mu,p_{\mathrm{T}}^\text{miss}) $. |
png pdf |
Table 2:
Target $ \tau_\mathrm{h} $ identification efficiencies for the different working points defined for the three discriminators. The target efficiencies are evaluated with the $ \mathrm{H}\to\tau\tau $ event sample for $ \tau_\mathrm{h} $ with $ p_{\mathrm{T}} \in $ [30, 70] GeV. |
png pdf |
Table A1:
Default values of the parameters used in the classification loss function for DEEPTAU training. |
| Summary |
| In this note, the newly deployed version of the DEEPTAU algorithm, v2.5, used to discriminate $ \tau_\mathrm{h} $ candidates against jets, electrons, and muons has been introduced. This deep convolutional neural network exhibits improved performance with respect to its predecessor, reducing the jet misidentification rate by 30-50% for a given reconstruction and identification $ \tau_\mathrm{h} $ efficiency. The implementation of domain adaptation by backpropagation, has reduced performance discrepancies between collision data and simulation, decreasing the necessary residual corrections. The domain adaptation was introduced by modifying gradient calculation in the neural network, via the addition of an adversarial subnetwork, designed to discriminate between collision data and simulations, in parallel to the tau classification task. The DEEPTAU algorithm, using both collision data and simulated samples, was therefore able to maximize the $ \tau_\mathrm{h} $ classification performance, while minimizing the data-simulation discrepancies. The algorithm was trained on simulated proton-proton collision data corresponding to the 2018 data-taking conditions, as well as data collected during the same year, which was used for domain adaptation. The DEEPTAU v2.5 discriminants were introduced in 2022 to be used by several CMS physics analyses. The algorithm has therefore been validated using both 2018 and 2022 collision data. The observed $ \tau_\mathrm{h} $ efficiencies were found to agree with the expected efficiencies from simulated events within 10% for 2018 and 15% for 2022. This agreement is improved with respect to the previous iteration of the algorithm and confirms the effectiveness of domain adaptation. |
| References | ||||
| 1 | CMS Collaboration | Observation of the Higgs boson decay to a pair of $ \tau $ leptons with the CMS detector | PLB 779 (2018) 283 | CMS-HIG-16-043 1708.00373 |
| 2 | CMS Collaboration | Search for Higgs boson pair production in events with two bottom quarks and two tau leptons in proton--proton collisions at $ \sqrt{s}= $ 13 TeV | PLB 778 (2018) 101 | CMS-HIG-17-002 1707.02909 |
| 3 | ATLAS Collaboration | Cross-section measurements of the Higgs boson decaying into a pair of $ \tau $-leptons in proton-proton collisions at $ \sqrt{s}= $ 13 TeV with the ATLAS detector | PRD 99 (2019) 072001 | 1811.08856 |
| 4 | ATLAS Collaboration | Search for resonant and nonresonant Higgs boson pair production in the $ {\text{b}\bar{\text{b}}\tau^+\tau^-} $ decay channel in pp collisions at $ \sqrt{s}= $ 13 TeV with the ATLAS detector | PRL 121 (2018) 191801 | 1808.00336 |
| 5 | ATLAS Collaboration | Test of CP invariance in vector-boson fusion production of the Higgs boson in the $ \text{H}\rightarrow\tau\tau $ channel in proton-proton collisions at $ \sqrt{s} = $ 13 TeV with the ATLAS detector | PLB 805 (2020) 135426 | 2002.05315 |
| 6 | CMS Collaboration | Search for additional neutral MSSM Higgs bosons in the $ \tau\tau $ final state in proton-proton collisions at $ \sqrt{s}= $ 13 TeV | JHEP 09 (2018) 007 | CMS-HIG-17-020 1803.06553 |
| 7 | ATLAS Collaboration | Search for charged Higgs bosons decaying via H$ ^{\pm} \to \tau^{\pm}\nu_{\tau} $ in the $ \tau $+jets and $ \tau $+lepton final states with 36 fb$ ^{-1} $ of pp collision data recorded at $ \sqrt{s} = $ 13 TeV with the ATLAS experiment | JHEP 09 (2018) 139 | 1807.07915 |
| 8 | CMS Collaboration | Search for an exotic decay of the Higgs boson to a pair of light pseudoscalars in the final state with two b quarks and two $ \tau $ leptons in proton-proton collisions at $ \sqrt{s}= $ 13 TeV | PLB 785 (2018) 462 | CMS-HIG-17-024 1805.10191 |
| 9 | CMS Collaboration | Search for a heavy pseudoscalar Higgs boson decaying into a 125 GeV Higgs boson and a Z boson in final states with two tau and two light leptons at $ \sqrt{s}= $ 13 TeV | JHEP 03 (2020) 065 | CMS-HIG-18-023 1910.11634 |
| 10 | CMS Collaboration | Search for lepton flavour violating decays of a neutral heavy Higgs boson to $ \mu\tau $ and e$ \tau $ in proton-proton collisions at $ \sqrt{s}= $ 13 TeV | JHEP 03 (2020) 103 | CMS-HIG-18-017 1911.10267 |
| 11 | CMS Collaboration | Search for a low-mass $ \tau^+\tau^- $ resonance in association with a bottom quark in proton-proton collisions at $ \sqrt{s}= $ 13 TeV | JHEP 05 (2019) 210 | CMS-HIG-17-014 1903.10228 |
| 12 | CMS Collaboration | Search for charged Higgs bosons in the H$ ^{\pm} \to \tau^{\pm}\nu_\tau $ decay channel in proton-proton collisions at $ \sqrt{s} = $ 13 TeV | JHEP 07 (2019) 142 | CMS-HIG-18-014 1903.04560 |
| 13 | ATLAS Collaboration | Search for heavy Higgs bosons decaying into two tau leptons with the ATLAS detector using pp collisions at $ \sqrt{s}= $ 13 TeV | PRL 125 (2020) 051801 | 2002.12223 |
| 14 | CMS Collaboration | Search for direct pair production of supersymmetric partners to the $ \tau $ lepton in proton-proton collisions at $ \sqrt{s}= $ 13 TeV | EPJC 80 (2020) 189 | CMS-SUS-18-006 1907.13179 |
| 15 | CMS Collaboration | Search for heavy neutrinos and third-generation leptoquarks in hadronic states of two $ \tau $ leptons and two jets in proton-proton collisions at $ \sqrt{s} = $ 13 TeV | JHEP 03 (2019) 170 | CMS-EXO-17-016 1811.00806 |
| 16 | ATLAS Collaboration | Searches for third-generation scalar leptoquarks in $ \sqrt{s} = $ 13 TeV pp collisions with the ATLAS detector | JHEP 06 (2019) 144 | 1902.08103 |
| 17 | CMS Collaboration | Analysis of the CP structure of the Yukawa coupling between the Higgs boson and $ \tau $ leptons in proton-proton collisions at $ \sqrt{s}= $ 13 TeV | JHEP 06 (2022) 012 | CMS-HIG-20-006 2110.04836 |
| 18 | CMS Collaboration | The CMS experiment at the CERN LHC | JINST 3 (2008) S08004 | |
| 19 | CMS Collaboration | Development of the CMS detector for the CERN LHC Run 3 | JINST 19 (2024) P05064 | CMS-PRF-21-001 2309.05466 |
| 20 | CMS Collaboration | Performance of $ \tau $-lepton reconstruction and identification in CMS | JINST 7 (2012) P01001 | CMS-TAU-11-001 1109.6034 |
| 21 | CMS Collaboration | Identification of hadronic tau lepton decays using a deep neural network | JINST 17 (2022) P07023 | CMS-TAU-20-001 2201.08458 |
| 22 | Y. Ganin and V. Lempitsky | Unsupervised domain adaptation by backpropagation | 1409.7495 | |
| 23 | CMS Collaboration | Performance of the CMS Level-1 trigger in proton-proton collisions at $ \sqrt{s} = $ 13 TeV | JINST 15 (2020) P10017 | CMS-TRG-17-001 2006.10165 |
| 24 | CMS Collaboration | The CMS trigger system | JINST 12 (2017) P01020 | CMS-TRG-12-001 1609.02366 |
| 25 | CMS Collaboration | Performance of the CMS high-level trigger during LHC run 2 | JINST 19 (2024) P11021 | CMS-TRG-19-001 2410.17038 |
| 26 | CMS Collaboration | Electron and photon reconstruction and identification with the CMS experiment at the CERN LHC | JINST 16 (2021) P05014 | CMS-EGM-17-001 2012.06888 |
| 27 | CMS Collaboration | Performance of the CMS muon detector and muon reconstruction with proton-proton collisions at $ \sqrt{s}= $ 13 TeV | JINST 13 (2018) P06015 | CMS-MUO-16-001 1804.04528 |
| 28 | CMS Collaboration | Description and performance of track and primary-vertex reconstruction with the cms tracker | JINST 9 (2014) P10009 | CMS-TRK-11-001 1405.6569 |
| 29 | J. Alwall et al. | The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations | JHEP 07 (2014) 079 | 1405.0301 |
| 30 | J. Alwall et al. | Comparative study of various algorithms for the merging of parton showers and matrix elements in hadronic collisions | EPJC 53 (2008) 473 | 0706.2569 |
| 31 | R. Frederix and S. Frixione | Merging meets matching in MC@NLO | JHEP 12 (2012) 061 | 1209.6215 |
| 32 | P. Nason | A new method for combining NLO QCD with shower Monte Carlo algorithms | JHEP 11 (2004) 040 | hep-ph/0409146 |
| 33 | S. Frixione, P. Nason, and C. Oleari | Matching NLO QCD computations with parton shower simulations: the POWHEG method | JHEP 11 (2007) 070 | 0709.2092 |
| 34 | S. Alioli, P. Nason, C. Oleari, and E. Re | A general framework for implementing NLO calculations in shower Monte Carlo programs: the POWHEG BOX | JHEP 06 (2010) 043 | 1002.2581 |
| 35 | S. Frixione, P. Nason, and G. Ridolfi | A positive-weight next-to-leading-order Monte Carlo for heavy flavour hadroproduction | JHEP 09 (2007) 126 | 0707.3088 |
| 36 | J. M. Campbell, R. K. Ellis, P. Nason, and E. Re | Top-pair production and decay at NLO matched with parton showers | JHEP 04 (2015) 114 | 1412.1828 |
| 37 | S. Alioli, P. Nason, C. Oleari, and E. Re | NLO single-top production matched with shower in POWHEG: $ s $- and $ t $-channel contributions | JHEP 09 (2009) 111 | 0907.4076 |
| 38 | E. Re | Single-top Wt-channel production matched with parton showers using the POWHEG method | EPJC 71 (2011) 1547 | 1009.2450 |
| 39 | T. Sjöstrand et al. | An introduction to PYTHIA 8.2 | Comput. Phys. Commun. 191 (2015) 159 | 1410.3012 |
| 40 | Y. Li and F. Petriello | Combining QCD and electroweak corrections to production in FEWZ | PRD 86 (2012) 094034 | 1208.5967 |
| 41 | M. Czakon and A. Mitov | Top++: A program for the calculation of the top-pair cross-section at hadron colliders | Comput. Phys. Commun. 185 (2014) 2930 | 1112.5675 |
| 42 | P. Kant et al. | HatHor for single top-quark production: Updated predictions and uncertainty estimates for single top-quark production in hadronic collisions | Comput. Phys. Commun. 191 (2015) 74 | 1406.4403 |
| 43 | CMS Collaboration | Event generator tunes obtained from underlying event and multiparton scattering measurements | EPJC 76 (2016) 155 | CMS-GEN-14-001 1512.00815 |
| 44 | N. Davidson et al. | Universal interface of TAUOLA: Technical and physics documentation | Comput. Phys. Commun. 183 (2012) 821 | 1002.0543 |
| 45 | GEANT4 Collaboration | GEANT4 --- a simulation toolkit | NIM A 506 (2003) 250 | |
| 46 | CMS Collaboration | Particle-flow reconstruction and global event description with the CMS detector | JINST 12 (2017) P10003 | CMS-PRF-14-001 1706.04965 |
| 47 | M. Cacciari, G. P. Salam, and G. Soyez | The $ \text{anti-k}_\text{t} $ jet clustering algorithm | JHEP 04 (2008) 063 | 0802.1189 |
| 48 | M. Cacciari, G. P. Salam, and G. Soyez | FastJet user manual | EPJC 72 (2012) 1896 | 1111.6097 |
| 49 | M. Cacciari and G. P. Salam | Dispelling the $ N^{3} $ myth for the $ k_{\mathrm{T}} $ jet-finder | PLB 641 (2006) 57 | hep-ph/0512210 |
| 50 | CMS Collaboration | Pileup mitigation at CMS in 13 TeV data | JINST 15 (2020) P09018 | CMS-JME-18-001 2003.00503 |
| 51 | CMS Collaboration | Jet energy scale and resolution in the CMS experiment in pp collisions at 8 TeV | JINST 12 (2017) P02014 | CMS-JME-13-004 1607.03663 |
| 52 | CMS Collaboration | Performance of missing transverse momentum reconstruction in proton-proton collisions at $ \sqrt{s} = $ 13 TeV using the CMS detector | JINST 14 (2019) P07004 | CMS-JME-17-001 1903.06078 |
| 53 | CMS Collaboration | ECAL 2016 refined calibration and Run2 summary plots | CMS Detector Performance Summary, CMS-DP-2020-021, 2020 CDS |
|
| 54 | CMS Collaboration | Performance of reconstruction and identification of $ \tau $ leptons decaying to hadrons and $ \nu_\tau $ in pp collisions at $ \sqrt{s}= $ 13 TeV | JINST 13 (2018) P10005 | CMS-TAU-16-003 1809.02816 |
| 55 | CMS Collaboration | Performance of electron reconstruction and selection with the CMS detector in proton-proton collisions at $ \sqrt{s} = $ 8 TeV | JINST 10 (2015) P06005 | CMS-EGM-13-001 1502.02701 |
| 56 | D. Bertolini, P. Harris, M. Low, and N. Tran | Pileup per particle identification | JHEP 10 (2014) 059 | 1407.6013 |
| 57 | S. Ioffe and C. Szegedy | Batch normalization: accelerating deep network training by reducing internal covariate shift | in Proceedings of the 32nd International Conference on International Conference on Machine Learning, ICML'15, 2015 | 1502.03167 |
| 58 | K. He, X. Zhang, S. Ren, and J. Sun | Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification | link | 1502.01852 |
| 59 | I. J. Goodfellow, Y. Bengio, and A. Courville | Deep Learning | MIT Press, Cambridge, MA, USA, 2016 link |
|
| 60 | T.-Y. Lin et al. | Focal loss for dense object detection | TPAMI 42 (2020) 318 | 1708.02002 |
| 61 | D. P. Kingma and J. Ba | Adam: A method for stochastic optimization | in Conference Track Proceedings, 3rd International Conference on Learning Representations, ICLR 2015 | 1412.6980 |
| 62 | T. Dozat | Incorporating Nesterov momentum into ADAM | in Conference Track Proceedings, 4th International Conference on Learning Representations, ICLR 2016 link |
|
| 63 | M. Abadi et al. | TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems | Software available from https://www.tensorflow.org/ |
|
| 64 | F. Chollet et al. | Keras | Software available from https://keras.io |
|
| 65 | CMS Collaboration | A deep neural network to search for new long-lived particles decaying to jets | MLST 1 (2020) 035012 | CMS-EXO-19-011 1912.12238 |
| 66 | L. Russell | Identification of Hadronic Tau Lepton Decays with Domain Adaptation using Adversarial Machine Learning at CMS | Master's thesis, Imperial College London, École Polytechnique Fédérale de Lausanne, 2022 | |
| 67 | CMS Collaboration | The CMS statistical analysis and combination tool: Combine | CSBS 8 (2024) 19 | CMS-CAT-23-001 2404.06614 |
| 68 | M. Baak, S. Gadatsch, R. Harrington, and W. Verkerke | Interpolation between multi-dimensional histograms using a new non-linear moment morphing method | NIM A 771 (2015) 39 | 1410.7388 |
| 69 | CMS Collaboration | Search for W' decaying to tau lepton and neutrino in proton-proton collisions at $ \sqrt{s} = $ 8 TeV | PLB 755 (2016) 196 | CMS-EXO-12-011 1508.04308 |
|
|
Compact Muon Solenoid LHC, CERN |
|
|
|
|
|
|