State-of-the-art Deep Learning and Artificial Intelligence

Trade Agreement: Canadian Free Trade Agreement (CFTA)
Tendering Procedures: All interested suppliers may submit a bid
Non-Competitive Procurement Strategy: Exclusive Rights
Comprehensive Land Claim Agreement: No
Vendor Name and Address: 
MILA - Institut québécois d'intelligence artificielle
6666 Saint-Urbain
Suite 200
Montréal Quebec
Canada
H2S3H1
Nature of Requirements: 

ADVANCE CONTRACT AWARD NOTICE (ACAN)
An ACAN is a public notice indicating to the supplier community that a department or agency intends to award a contract for goods, services or construction to a pre-identified supplier, thereby allowing other suppliers to signal their interest in bidding, by submitting a statement of capabilities. If no supplier submits a statement of capabilities that meets the requirements set out in the ACAN, on or before the closing date stated in the ACAN, the contracting officer may then proceed with the award to the pre-identified supplierDefinition of the requirement 
The Department of NRCan has a requirement to conduct a research to provide mineral industry with the next generation of geoscience knowledge and innovative techniques. An important component of this research involves the processing and interpretation of 3D reflection seismic data and their integration with geochemical and geological data for producing accurate 3D representations of the subsurface. These 3D models improve targeting of deeply buried mineral deposits and for assessing mineral potential in areas that are covered by glacial sediments. The work will involve the following: The data required for this research will be provided by Natural Resources Canada and will include a combination of 3D seismic data (poststack and prestack), wireline logs, 3D geological models, geological logs from drill holes, rock geochemistry, and additional geophysical data from several mining camps across Canada. Note that the availability of type of data described above varies greatly between mining sites. Datasets that will be provided are representative of datasets typically used in mineral exploration. As such, they are heterogeneous, unbalanced in terms of class distribution, densely sampled in some areas but sparse in other areas, collected at various scales, subject to interpretation and pre-processing with possible associated artifacts, and have high uncertainty especially near edges of 3D geophysical and geological representations.

The main objective of this project is to integrate Deep Learning and AI algorithms in Natural Resources Canada’s subsurface modelling workflow that will contribute to improving geobody interface detection and volumetric segmentations, faults, and rock type prediction including alteration zones and mineralized bodies using seismic, geology, and other support data. Any parts of the conventional workflow including the data processing, 3D geological modeling, and their integration can be assessed and evaluated as long as benefits strongly contributes to the main objective of the project. The application of these algorithms for integrating other geophysical, geochemical and geological datasets to improve mineral potential modelling should also be considered.

1. GENERAL TASKS
From a general perspective, the project will follow a rigorous scientific methodology for technical experiments listed in section “Specific Tasks” ensuring that results and findings are sound. The general methodology consists of the following steps that will be applied whenever appropriate:

1.1 Problem formalization: a sound problem definition will be required for each idea / experiment. The problem formalization will also be done for each new idea or opportunity identified during the project.

1.2 Literature review: all experiments have relevant literature which will help guide model and architecture choices. Some simpler experiments such as the automatic first break picking are foreseen to use well known approaches which will probably require less effort for the literature review while some other experiments are based on very recent and more specialized work thus requiring more effort.

1.3 Experimental protocol definition: we will define a protocol which will structure each experiment, specifying the following elements:
 Metrics: choose the performance evaluation metrics that will be used to train and evaluate models.
 Evaluation protocol: define the partitioning of the data into train, validation and test datasets. It will be of the utmost importance to carefully design the splitting strategy to mitigate the effects of spatial correlation, and the effect associated with the underlying imprecision of the workflow used to produce the datasets. The splitting strategy will be carefully crafted to provide insight about model generalization capabilities.

1.4 Algorithm implementation and optimization: models will be implemented using PyTorch 1.6 or later following development best practices. Each implemented model will make optimal usage of all available computational resources.

1.5 Model training and result collection: once a stable algorithm implementation is reached, model training will proceed through a hyperparameter search.

1.6 Result analysis and presentation: results will be carefully analyzed and presented to both Natural Resources Canada and Mila teams, with several goals in mind:
 Understand the weaknesses and strengths of the models, taking into consideration the quality and quantity of the data, the differences in the characteristics of the training, validation and test partitions, and the generalization capabilities of the models.
 Identify opportunities for improving the models (leading to a new round of experiments).

1.7 Identify new ideas that are worth exploring.

1.8 Review the prioritization of ideas and of next steps.

2. SPECIFIC TASKS
The tasks in this section are associated with tangible outputs that form the deliverables of this work. It is understood that the methodology defined in the general tasks will be applied where appropriate to these tasks.

2.1 First break picking: Train a predictive machine learning model that detects the onset of arrivals of seismic signals on the receivers. The deep learning model will consider the signal from a group of receivers (i.e. shot gather or receiver lines) instead of the classical trace-by-trace approach. A set of several thousand shot gathers, originating from different sites will be used to validate the generalization capability of the trained models. A significant portion of these shot gathers will be manually annotated and considered as “ground truth”. Labels generated automatically by the algorithm used by Natural Resource Canada will serve to create a performance baseline.

2.2 Seismic volumes data loading and analysis: Data curation and data visualization aimed at developing a good understanding of the data, which will in turn help to build better models and facilitate the interpretation of the results. A data processing pipeline will be built to prepare all datasets in respect of the experimental protocols. Since many of the experiments will use the same data, the data processing pipeline will be thoughtfully designed and reused in all experiments.

2.3 Supervised learning of 3D reflection seismic data: train predictive machine learning models using existing geological labels and additional support data provided by Natural Resources Canada. Initial development will be done using synthetic data for which all parameters are known. The approach will be tested on field data (seismic and geology) provided by Natural Resources Canada. A supervised learning framework will be used for rock-type classification at the voxel level and boundary prediction between different rock types. The latter consists in identifying the reflection points over the signal where an important interface boundary occurs.

2.4 Self-supervised learning of 3D reflection seismic data: Leverage unlabeled data to learn useful representations for voxel classification and boundary detection with self-supervised learning. Define a “pretext” task for which no labeled data is required and train a non-linear encoder from scratch on this pretext task alone, and then fine-tune with the actual labels from the main tasks. Self-supervised learning is a rapidly evolving field in the deep learning community. A literature review will need to be done at the beginning of the project.

2.5 Article preparation: prepare a draft of a manuscript for publication in peer-reviewed scientific journal. The most promising element(s) of this work will form the basis of the manuscript and will be determined near the end of the project.

3. DELIVERABLES
A table detailing the specific project tasks, their level of effort in person-days, and the deliverables they contribute to will be provided on request.

3.1 Criteria for assessment of the Statement of Capabilities (Minimum Essential Requirements) 

 Any interested supplier must demonstrate by way of a statement of capabilities that it meets the following requirements: 

 Experience in conducting state-of-the-art academic research on Deep Learning using supervised, semi-supervised, unsupervised approaches with spatial and preferably three-dimensional spatial data. Such experience must be demonstrated by providing a list of peer-reviewed publications on appropriate topics in scientific journals.

 Academic qualifications: Must have a research team with all members having a Ph.D. in the field of machine learning or artificial intelligence.

 In-depth knowledge of three-dimensional geological and reflection seismic data and their integration for the prediction of mineral resources. Knowledge of data processing sequences (for both geology and seismic data) and their impacts on final results and possible errors is also required. Knowledge can be demonstrated by providing scientific publications, report, project proposals combining both geology and seismic data.


4. APPLICABILITY OF THE TRADE AGREEMENT(S) TO THE PROCUREMENT 
This procurement is subject to the following trade agreement(s) 
 Canadian Free Trade Agreement (CFTA)

5. JUSTIFICATION FOR THE PRE-IDENTIFIED SUPPLIER 

Mila (https://mila.quebec/en/) is a non-profit research institute in artificial intelligence which rallies 500 researchers specializing in the field of deep learning. Mila’s mission is to be a global pole for scientific advances that inspires innovation and the development of AI for the benefit of all. Since 2017, Mila is the result of a partnership between the Université de Montréal and McGill University, closely linked with Polytechnique Montréal and HEC Montréal, and gathers in its offices a community of professors, students, industrial partners and startups working in AI, making the institute the world’s largest academic research center in deep learning.

Of the three National research institutes specialized in AI and Deep learning (i.e. Vector Institute, Toronto, and Amii, Edmonton), Mila is the only one that has developed an in-depth knowledge and expertise in Deep Learning algorithms and architectures suitable for mineral exploration conducted with a combination of geological and geophysical data. The development of AI and Deep learning approaches for mineral exploration has not been considered as field of interest by the two other research institutes. Thus, Mila is the only research institute on Deep Learning and AI able to perform this work. 

NRCan and Mila have signed an industrial partnership agreement to develop and apply artificial intelligence and machine learning research in the field of natural resources. Following this partnership agreement, NRCan has awarded a contract to Mila to design a research-oriented roadmap for the development of novel Deep learning architectures for the generation of 3D models to improve our understanding of the subsurface and detect new deeply-buried  mineralized bodies. The work proposed in this procurement is the implementation of this roadmap to develop new tools to improve the efficiency of mineral exploration. 

In brief, Mila is the only pre-identified supplier that can undertake this leading edge research in artificial intelligence for geological characterization.

6.GOVERNMENT CONTRACTS REGULATIONS EXCEPTION(S) 
The following exception(s) to the Government Contracts Regulations is invoked for this procurement under subsection 6(d) - "only one person is capable of performing the work").  we confirm that only one organization is capable of performing the proposed work.  Mila institute is the only AI institution that has undertaken geological characterization with geophysical and seismic datasets.  In particular, Mila is not only applying complex AI algorithms to our geoscientific questions, but they are developing leading edge methodology and developing fundamental research in AI.

7. OWNERSHIP OF INTELLECTUAL PROPERTY 
Ownership of any Foreground Intellectual Property arising out of the proposed contract will vest in the Contractor.

8. PERIOD OF THE PROPOSED CONTRACT OR DELIVERY DATE : 
The proposed contract is for a period of less than 2 year, from January 11, 2021 to March 31, 2022. 

9. NAME AND ADDRESS OF THE PRE-IDENTIFIED SUPPLIER :
Mila-Quebec AI Institute
6666 St Urbain Street
Montreal, Quebec H2S 3H1

10.SUPPLIERS' RIGHT TO SUBMIT A STATEMENT OF CAPABILITIES 
Suppliers who consider themselves fully qualified and available to provide the goods, services or construction services described in the ACAN may submit a statement of capabilities in writing to the contact person identified in this notice on or before the closing date of this notice. The statement of capabilities must clearly demonstrate how the supplier meets the advertised requirements.

11.CLOSING DATE FOR A SUBMISSION OF A STATEMENT OF CAPABILITIES 
The closing date and time for accepting statements of capabilities is December 31, 2020 at 2:00 p.m. EST. 

Inquiries and submission of statements of capabilities Inquiries and statements of capabilities are to be directed to: 

Annick Monfette, Spécialiste en approvisionnements/ Supply Specialist , 
Direction des achats innovateurs (DAI)/ Innovation Procurement Directorate (IPD)
Services Publics et Approvisionnement Canada (SPAC) / gouvernement du Canada
Public Services and Procurement Canada (PSPC) / Government of Canada?? annick.monfette@tpsgc-pwgsc.gc.ca ? 873-355-1907

Delivery Date: Above-mentioned

You are hereby notified that the government intends to negotiate with one firm only as identified above. Should you have any questions concerning this requirement, contact the contracting officer identified above.

An Advance Contract Award Notice (ACAN) allows departments and agencies to post a notice, for no less than fifteen (15) calendar days, indicating to the supplier community that it intends to award a good, service or construction contract to a pre-identified contractor. If no other supplier submits, on or before the closing date, a Statement of Capabilities that meets the requirements set out in the ACAN, the contracting authority may then proceed with the award.  However, should a Statement of Capabilities be found to meet the requirements set out in the ACAN, then the contracting authority will proceed to a full tendering process.

Suppliers who consider themselves fully qualified and available to provide the services/goods described herein, may submit a statement of capabilities in writing to the contact person identified in this Notice on or before the closing date of this Notice. The statement of capabilities must clearly demonstrate how the supplier meets the advertised requirements.

The PWGSC file number, the contracting officer's name and the closing date of the ACAN must appear on the outside of the envelope in block letters or, in the case of a facsimile transmission, on the covering page.

The Crown retains the right to negotiate with suppliers on any procurement.

Documents may be submitted in either official language of Canada.